Intelligent control method for lead-bismuth alloy smelting and refining process

Abstract

This invention discloses an intelligent control method for the smelting and refining process of lead-bismuth alloys, comprising: A1: acquiring visual images of alkaline slag from the alkaline refining molten pool of lead-bismuth alloys, molten pool smelting temperature data, and stirring operation data, and preprocessing them respectively; A2: acquiring output features from the dust stripping branch, thermal distortion stripping branch, and color deviation stripping branch; A3: performing adaptive fusion to sequentially calculate the characteristic images of alkaline slag from the pure molten pool; A4: calculating the effective alkaline slag visual features; A5: calculating the reaction progress and the critical risk value of excessive bismuth metal burn-off; A6: calculating the reaction termination judgment flag, and terminating the smelting reaction when the reaction termination judgment flag is 1. This invention can solve the problems of traditional methods being difficult to adapt to complex fields of view with strong interference, unable to accurately identify the reaction progress and the critical point of burn-off, and thus difficult to balance the impurity removal effect and the control of bismuth metal oxidation burn-off.

Description

Technical Field

[0001] This invention relates to the field of lead-bismuth alloy refining technology, and in particular to an intelligent control method for the lead-bismuth alloy smelting and refining process. Background Technology

[0002] Eutectic lead-bismuth alloys are key functional materials in advanced nuclear energy system cooling, high-end electronic packaging, and special metallurgy. The uniformity of their composition and the precision of removing harmful impurities such as tin, arsenic, and antimony are core indicators determining the application performance and quality grade of the product. Oxidation refining and alkaline refining of tin, arsenic, and antimony are the core processes in lead-bismuth alloy smelting. This process is typically carried out in a high-temperature alkaline environment of 450-500℃. Impurities are removed by adding alkaline refining agents in batches and coordinating with strong stirring of the molten pool. During this process, it is necessary to control the oxidation and burn-off of metallic bismuth at high temperatures while ensuring the effectiveness of impurity removal. The control requirements for the reaction state inside the furnace are far higher than in conventional smelting processes.

[0003] Currently, the control of this refining process in the industry mainly relies on the manual experience of operators. During the operation, alkaline volatile fumes and dust are continuously generated on the surface of the molten pool. Batch feeding and strong stirring operations further amplify the interference in the field. Under these conditions, operators need to judge the reaction progress of each batch by observing the color and slag state changes of the alkaline slag, thereby determining the timing of adding refining agents and the reaction endpoint. At the same time, the strong thermal convection of the molten pool under high temperature conditions will cause optical distortion, resulting in distorted and shaky images. The high-temperature thermal radiation of the melt itself will superimpose on visible light to produce color deviation, obscuring the true subtle color and state changes of the alkaline slag, leading to large errors in manual judgment and easily causing problems such as inadequate impurity removal or excessive burning of bismuth metal. Traditional offline sampling and testing methods have detection lag and cannot meet the needs of dynamic control of the refining process.

[0004] In recent years, although deep learning-based machine vision technology has begun to be applied in the field of metallurgical identification, it still has significant limitations in extreme working conditions such as lead-bismuth alloy refining. Existing technologies typically employ general image enhancement or global denoising algorithms. These single strategies are insufficient to simultaneously address the combined interference of three distinctly different physical characteristics: smoke and dust occlusion, thermal distortion, and thermal radiation color shift. While eliminating interference, general algorithms are prone to smoothing or blurring key texture features such as the fine granularity of alkali slag edges, the looseness of slag accumulation, and the surface gas barrier layer. These features are core indicators for judging the adequacy of impurity removal.

[0005] Furthermore, existing solutions often disconnect visual information from process conditions. In dynamic refining processes, the appearance evolution of the alkali residue does not always correspond to the actual reaction progress. For example, instantaneous changes in stirring intensity or temperature fluctuations can cause false visual feedback from the alkali residue, triggering numerous false positive interference signals. Due to the lack of a deep integration mechanism that combines visual characteristics with temperature, stirring conditions, and chemical reaction kinetics (such as Arrhenius's law), existing technologies struggle to accurately identify the critical points of impurity removal progress and bismuth oxidation loss from multiple dimensions. This results in the system's inability to achieve high-confidence adaptive control and feedforward early warning in complex and variable refining environments, making it difficult to effectively balance the contradiction between impurity removal efficiency and metal loss control. Summary of the Invention

[0006] In view of this, the present invention aims to provide an intelligent control method for the smelting and refining process of lead-bismuth alloys, so as to solve the problems that traditional methods are difficult to adapt to strong interference and complex field of view, cannot accurately identify the reaction progress and the burn-off critical point, and thus are difficult to balance the impurity removal effect and the control of bismuth metal oxidation burn-off.

[0007] A method for intelligent control of lead-bismuth alloy smelting and refining processes includes: A1: Collect visual images of alkaline slag, melting temperature data, and stirring operation data of the alkaline refining pool of lead-bismuth alloy, and preprocess them respectively to obtain preprocessed visual images of alkaline slag, preprocessed melting time-series temperature data, and preprocessed stirring operation sequence data. A2: Based on the pre-processed visual image of the alkali slag, the shallow general features of the furnace image, the probability map of smoke and dust obstruction, the probability map of thermal distortion, and the probability map of color deviation residue are extracted in sequence. Then, through the three-channel parallel stripping mechanism of molten pool interference decoupling, the output features of the smoke and dust stripping branch, the output features of the thermal distortion stripping branch, and the output features of the color deviation stripping branch are extracted. A3: Based on the output characteristics of the dust stripping branch, the thermal distortion stripping branch, the color deviation stripping branch, as well as the dust occlusion probability map, the thermal distortion probability map, the color deviation residual probability map, and the shallow general features of the furnace image, adaptive fusion is performed to calculate the preliminary interference stripping features, the alkaline slag texture protection soft mask, and the pure molten pool alkaline slag feature image in sequence. A4: Based on the characteristic image of alkaline slag in the pure molten pool, extract the color gradient feature, thin-thickness texture feature, and bismuth oxidation burn-off feature of the alkaline slag in sequence, and calculate the interference residual risk coefficient map, the visual features of the fused alkaline slag, and the visual features of the effective alkaline slag. A5: Based on the effective visual characteristics of alkali slag, the molten pool smelting time-series temperature data after pretreatment, and the stirring operation sequence data after pretreatment, calculate the reaction advancement degree and the critical risk value of excessive bismuth metal burn-off. A6: Based on the reaction progress, the critical risk value of excessive bismuth metal burn-off, and the smelting time sequence temperature data of the pretreated molten pool, evaluate the termination conditions of the refining reaction, obtain the reaction termination judgment mark, and terminate the smelting reaction when the reaction termination judgment mark is 1.

[0008] Furthermore, step A1 includes: A11: Visual images of alkaline slag in the alkaline refining pool of lead-bismuth alloy were acquired by an industrial high-temperature dustproof camera device. Then, Gaussian filtering for noise reduction, optical distortion correction, and thermal radiation color shift compensation were applied sequentially to obtain the pre-processed visual images of alkaline slag. A12: The melting temperature data of the lead-bismuth alloy refining furnace is collected by armored thermocouples. The data type is one-dimensional time-series numerical data. The melting temperature data of the melting pool is processed by removing three times the standard difference constant and completing the missing linear interpolation to obtain the preliminary pre-processed melting time-series temperature data of the melting pool. A13: The stirring operation data of the refining furnace stirring actuator is collected through the stirring mechanism condition acquisition module. The data type is one-dimensional time-series numerical data, including the real-time stirring speed, stirring start and stop status, and stirring operation sequence cycle. The stirring operation data is processed by three times standard difference constant value removal and time sequence normalization to obtain the preliminary pre-processed stirring operation sequence data. A14: The visual images of the pre-treated alkali slag, the molten pool smelting time-series temperature data, and the stirring operation sequence data after the preliminary pretreatment are synchronized with timestamps. Linear interpolation or spline interpolation methods are used to resample all data to a uniform sampling frequency to obtain the visual images of the pre-treated alkali slag, the molten pool smelting time-series temperature data, and the stirring operation sequence data after the preliminary pretreatment.

[0009] Furthermore, step A2 includes: A21: Based on the preprocessed visual image of the alkali slag, shallow general features of the furnace image are extracted using a convolutional encoder with residual connections. The probability maps of smoke occlusion, thermal distortion, and color cast residue at each pixel location are calculated using point-by-point convolution and channel segmentation. The calculation method is as follows: ; ; ; ; ; in, This is a common feature of shallow layers in the furnace surface. For convolutional layers with residual connections, Visual images of the pretreated alkali residue. The spatial gradient features are general shallow features of the furnace image. For the Sobel edge gradient operator, The fusion features after splicing This is for channel splicing operations. To fuse the interference probability tensor, For the Softmax function, For pointwise convolutional layers, This is an operation to divide the channel equally along its dimension. A probability diagram showing the obstruction caused by smoke and dust. This is a thermal distortion probability map. This is a probability diagram of color deviation residue. A22: Based on the shallow general features of the furnace image and the probability maps of smoke and dust obstruction, thermal distortion, and color deviation residue, a three-channel parallel stripping mechanism for decoupling molten pool interference is used to strip away smoke and dust obstruction, thermal convection distortion, and color deviation residue respectively, resulting in the output features of the smoke and dust stripping branch, the thermal distortion stripping branch, and the color deviation stripping branch. The calculation method is as follows: ; ; ; in, The output characteristics of the dust stripping branch are as follows. For element-wise subtraction, For dimensional expansion operations, For Hadama accumulation, For dilated convolution operators, This represents the output characteristics of the thermal distortion stripping branch. It is a convolutional layer. The output features of the color-bias stripping branch are as follows: This is a channel attention mechanism.

[0010] It should be further explained that in a high-temperature alkaline environment, the surface of the molten pool is not clearly visible; strong alkaline fumes will directly obscure the view, making it impossible for operators and visual sensors to accurately observe the state of the alkaline slag inside the molten pool; the heat convection above the high-temperature molten pool causes the image to shake and distort, destroying the geometric features of the alkaline slag outline; at the same time, the high-temperature heat radiation of the molten metal itself produces a strong red hue, and this color deviation masks the true color change of the alkaline slag during the impurity removal process, making it difficult to judge the progress of impurity removal by color. This invention establishes an interference decoupling and stripping mechanism for the molten pool environment in stage A2. First, feature extraction is performed on the preprocessed alkaline slag image, transforming the original pixel information into higher-order feature representations. Simultaneously, the spatial gradient information of the features is calculated—gradients can capture the distribution of edges and textures, helping to distinguish interference areas from effective information areas. Second, the basic features and gradient features are fused to calculate the probability maps of smoke occlusion, thermal distortion, and color cast residue. These three probability maps quantify the degree of contamination by three types of interference in the image with pixel-level precision, providing location information for the location and intensity of the interference. Third, based on these three probability maps, three parallel stripping branches are designed. The smoke stripping branch uses air... Hole convolution adaptively transforms the basic features, and then uses the smoke and dust probability map to suppress false features, gradually eliminating interference information in the smoke and dust-covered areas. The thermal distortion stripping branch processes the thermal distortion probability map through convolution, correcting the geometric distortion in the basic features caused by thermal convection and restoring the outline and shape of the alkali slag. The color deviation stripping branch uses a channel mechanism to reweight different color channels of the basic features, modulates the color information distorted by thermal radiation red light through the color deviation probability map, and highlights the true color of the alkali slag. After interference stripping by A2, subsequent alkali slag feature extraction and process decision can be based on more realistic and reliable visual input, thereby improving the system's accuracy in recognizing the state of the molten pool alkali slag. Traditional image processing methods typically employ single global enhancement or general denoising algorithms, which are often limited in their effectiveness against complex interferences in the molten pool environment. A single strategy struggles to address the three types of interference with different physical properties: smoke and dust obscuration, thermal distortion, and color cast. As a result, while some interferences are weakened, they are not truly eliminated, and the characteristic information of the alkali slag remains partially contaminated, affecting subsequent feature recognition and process decisions. General enhancement methods, while removing interference, often weaken or distort the true characteristics of the alkali slag—fine slag particle texture and edge morphology information are easily smoothed or blurred. These features are key indicators for judging the adequacy of impurity removal; losing this information leads to inaccurate process status judgments. This invention, through a multi-channel parallel design, ensures a targeted stripping mechanism for each type of interference, enabling targeted elimination of smoke, thermal distortion, and color cast. This avoids incomplete elimination of certain types of interference and the accidental deletion of useful features due to overprocessing.

[0011] Furthermore, step A3 includes: A31: Based on the output characteristics of the smoke and dust stripping branch, the output characteristics of the thermal distortion stripping branch, the output characteristics of the color shift stripping branch, the smoke and dust occlusion probability map, the thermal distortion probability map, and the color shift residue probability map, the three stripping results are adaptively fused to obtain preliminary interference stripping characteristics. The calculation method is as follows: ; ; ; in, To fuse the weight tensor for interference coupling, This is a convolutional layer with a kernel size of 3×3. Spatial fusion weight map of smoke and dust stripping branches. This is a spatial fusion weight map of the thermal distortion stripping branch. For the spatial fusion weight map of the color bias stripping branch, This is an operation to divide the channel equally along its dimension. Preliminary interference stripping characteristics, This is element-wise addition; A32: Based on the initial interference stripping features and the shallow general features of the furnace image, an alkaline slag texture protection soft mask is calculated using an alkaline slag texture fidelity adaptive mask refinement mechanism. The initial interference stripping features are then refined to obtain a pure molten pool alkaline slag feature image. The calculation method is as follows: ; ; ; ; in, The spatial gradient features are the initial interference stripping features. To standardize gradient similarity, It is a very small positive number. A soft mask is used to protect the texture of alkali residue. For the Sigmoid function, This is the gradient similarity amplification factor. This is a characteristic image of alkaline slag from a pure molten pool.

[0012] It should be further explained that although the three-channel parallel stripping in the A2 stage can eliminate the three types of interference—smoke, thermal distortion, and color deviation—a problem still exists in the actual process: the output features of the three branches are processed independently, and how to reasonably fuse them into unified visual data, and how to protect the subtle features of the alkali slag during the fusion process; especially during the heating and isothermal periods of the molten pool, the degree of looseness of the alkali slag accumulation, the thickness and connectivity of the surface gas barrier layer, and other detailed features are key indicators for judging the adequacy of impurity removal. These features often manifest as edge morphology and texture changes; if these details are accidentally smoothed or blurred during the interference elimination process, even if the interference is completely removed later, the loss of alkali slag features will lead to inaccurate process judgment. This invention employs an adaptive fusion mechanism in stage A3 to address this issue. First, in step A31, the probability maps output from the three branches are fused, and the interference coupling degree fusion weight tensor is calculated. This tensor reflects the relative intensity of the three types of interference at different locations in the image with pixel-level precision—in some areas, smoke and dust may have the greatest impact, while in others, thermal distortion may be the primary interference. Based on this weight distribution, the output features of the three branches are spatially weighted and fused, ensuring that each pixel location receives the most suitable interference removal result, rather than applying a uniform fusion strategy to all regions. This results in preliminary interference stripping features that include the results of smoke and dust removal, as well as thermal distortion correction and color shift modulation—a comprehensive output of three independent processes. Then, in step A32, to prevent the subtle features of the alkali residue from being damaged during the fusion process, a texture-preserving mask mechanism is introduced. This mechanism calculates the preliminary interference stripping features... Spatial gradients are compared with the gradients of the original base features. Regions with high gradient similarity indicate that the edge and texture features of the alkali slag are well preserved, while regions with low gradient similarity indicate that the detailed features may have been affected during the interference removal process. Based on this pixel-by-pixel similarity evaluation, an adaptive soft mask is generated. This mask tends to trust the interference stripping results in regions with high gradient similarity, while retaining more detailed information from the original base features in regions with low similarity. This approach utilizes the results of interference removal while also protecting the original texture of the alkali slag when necessary. Finally, the original features and stripped features are adaptively blended using the mask to obtain a clean molten pool alkali slag feature image. In areas where interference has been effectively removed and alkali slag features are preserved, the stripping results are used. In areas where interference removal may have damaged the features, the original features are preserved or supplemented, thus ensuring that the final output has neither significant interference contamination nor loss of key alkali slag features. Compared to existing technologies, traditional multi-channel fusion typically employs fixed weights or simple weighted averaging, failing to dynamically adjust based on local image characteristics. The fusion strategy is identical in high-interference and low-interference regions, easily leading to incomplete interference removal in some areas or unnecessary modification of features in regions with already weak interference. This invention, through interference coupling degree weighted fusion, enables spatial adaptability in the fusion process, allowing for differentiated processing based on the interference intensity distribution in different regions, thus improving fusion effectiveness. Furthermore, this invention introduces a alkali slag texture fidelity mask to directly detect whether key features of the alkali slag itself—edges, texture, and graininess—are preserved. This design ensures that the final output visual data is both less interfered with and retains the complete alkali slag features, providing a high-fidelity visual foundation for subsequent impurity removal degree judgment and process timing control.

[0013] Furthermore, step A4 includes: A41: Based on the characteristic image of alkaline slag from a pure molten pool, the color gradient features, thinning / dense texture features, and bismuth oxidation burn-off features of the alkaline slag are extracted using a multi-scale collaborative gated feature fusion mechanism. The calculation method is as follows: ; ; ; in, The color of the alkali residue is characterized by a gradual change. For residual blocks with residual connections, For hollow convolution, It exhibits the characteristic thin-to-thick texture of alkali residue. Characterized by bismuth oxidation and burn-off; A42: Based on the color gradient characteristics, viscosity texture characteristics, and bismuth oxidation burn-off characteristics of the alkali slag, the interference residue risk coefficient at each spatial location is assessed, resulting in an interference residue risk coefficient map. The visual characteristics of the fused alkali slag are then calculated using the following method: ; ; ; ; in, For the residual risk coefficient diagram of interference, These are the color component, texture component, and burn-off component, respectively. This is a feature that mitigates the risk of interference. Visual characteristics of the fused alkali residue; A43: Based on the fused visual characteristics of the alkali residue, weighted fusion is performed through collaborative gating to obtain the effective visual characteristics of the alkali residue. The calculation method is as follows: ; ; in, To coordinate gating weights, It is a multilayer perceptron. For global average pooling, The visual characteristics of the fused alkali residue To effectively identify the visual characteristics of alkali residue, This is a residual block with residual connections.

[0014] It should be further explained that after interference elimination and texture fidelity processing using A2 and A3, the visual data is relatively pure, but challenges still exist in key aspects of process decision-making. Process decisions for alkaline refining rely on observation of three dimensions of the molten pool slag: first, the color gradation of the slag, with different hues at different stages, and color evolution indicating the progress of impurity removal; second, the viscosity of the slag, with thin slag often indicating that the impurity removal reaction is nearing completion, while thick slag indicates that the reaction is still ongoing; and third, whether there is excessive oxidation and burn-off of bismuth metal. However, in actual production, these three characteristics can influence each other. For example, color changes may originate from both impurity removal and bismuth metal oxidation loss—operators find it difficult to accurately distinguish with the naked eye whether the color change reflects the progress of impurity removal or the beginning of significant bismuth loss; the judgment of viscosity is also affected by residual heat radiation, potentially leading to misjudgments of reaction sufficiency. In the A4 stage, this invention employs a multi-dimensional collaborative feature extraction and risk assessment mechanism. Firstly, in step A41, based on the characteristic image of the pure molten pool alkali slag, three independent feature extraction branches are specifically set up. The color gradient feature extraction branch uses a standard convolutional layer combined with a residual structure to focus on capturing the spatial color distribution and gradient trend of the alkali slag, reflecting the details of color evolution during impurity removal. The sludge-density texture feature extraction branch uses dilated convolution to capture the texture and structural features of the alkali slag within a large receptive field, especially visual indicators related to sludge density such as particle size and fluidity. The bismuth oxidation burn-off feature extraction branch uses pointwise convolution to capture local pixel-level information related to metal oxidation characteristics. In step A42, this invention introduces an interference residue risk assessment mechanism. Although the interference has been significantly weakened by the previous processing, residual thermal radiation or other weak interference may still remain in some locations, which may affect the accuracy of the features. Therefore, based on the feature outputs of the three branches, an interference residue risk coefficient map is calculated using convolution and the Sigmoid function. This map assesses the threat level of residual interference to feature credibility at different locations in the image with pixel-level accuracy—the risk coefficient is higher in locations with more residual interference and lower in locations where the interference has been fully cleared. Subsequently, the features of the three branches are weighted and weakened using the risk coefficient. The weight of the features in locations with high risk is reduced, which is equivalent to marking areas with low information credibility at the feature level, preventing the confusing information in these areas from misleading subsequent decisions. Then, the weakened features are fused, and adaptive gating is used to further enhance the contribution of areas with high feature strength and low interference risk, thereby obtaining a fused feature that integrates the three dimensions of process indicators and filters out the influence of interference residue. In step A43, the present invention performs collaborative gating weighting on the fused features; the gating mechanism extracts global statistical information of the features through global average pooling, and then uses a multilayer perceptron to calculate a scalar weight, which reflects the quality of the current fused feature - indicating how much the feature can represent the real alkali slag reaction state; based on this weight, the fused feature and its residual enhanced version are then weighted and combined to obtain the final effective alkali slag visual features; Compared to existing technologies that typically extract all information from a unified feature space, failing to effectively separate multiple coupled process indicators and resulting in the mixing of color, texture, and burn-off features, subsequent decision models struggle to accurately identify which type of process progress the current visual changes represent. This invention achieves multi-dimensional decoupled feature extraction by setting independent feature extraction pathways for each of the three process judgment dimensions—the color gradient branch emphasizes color information, the sparse-dense texture branch emphasizes structure and graininess, and the burn-off feature branch emphasizes local markings of metal oxidation. This allows for the acquisition of purified versions of features even when the three dimensions are correlated. Furthermore, an interference residue risk assessment mechanism further filters interference at the feature level, preventing deviations in process judgment due to residual interference and enabling the model to more accurately understand the true process state of the current molten pool.

[0015] Furthermore, step A5 includes: First, the pre-processed molten pool melting time-series temperature data is processed by a gated loop unit and then by a linear layer; the pre-processed stirring running time-series condition data is processed by a one-dimensional convolutional network and then by a linear layer; the above two processing results are spliced ​​together and then processed by a multilayer perceptron to obtain the process condition characteristics. The effective alkali residue visual features are processed by a convolutional layer with a kernel size of 1×1, and then the outer product operation is performed with the process condition features. After processing by the Sigmoid function, the process-visual affinity weight map is obtained. The process-visual affinity weight map is then multiplied by the Hadamard product with the effective alkali residue visual features to obtain the condition-aligned visual features. The visual features aligned with the working conditions are added element by element to the expanded process condition features, and then processed sequentially by a convolutional layer with a kernel size of 3×3 and a ReLU function to obtain the spatial distribution matrix of the reaction intensity. The spatial distribution matrix of the reaction intensity is then processed by global average pooling, and then processed sequentially by a multilayer perceptron and a Sigmoid function to obtain the degree of reaction advancement. The critical activation level for burn-off is obtained by subtracting the product of the material burn-off coefficient and the exponential function from the reaction advancement rate, followed by processing with the Sigmoid function; where the power of the exponential function is: ; in, For activation energy constraint parameters, To retrieve the last value, This refers to the pre-processed molten pool smelting timing temperature data; The critical activation value for bismuth burn-off is obtained by concatenating the results of global average pooling processing with the spatial distribution matrix of reaction intensity, followed by processing by multilayer perceptron and sigmoid function.

[0016] It should be further explained that in the dynamic environment of high-temperature smelting, changes in the appearance of alkali slag do not always reflect the actual progress of impurity removal. For example, a sudden increase in stirring speed may cause unreacted melt at the bottom to roll up, making the alkali slag appear to have changed color, but this change is actually just a momentary effect of the stirring conditions. Or, if the temperature fluctuates in a short period of time, the color and texture of the alkali slag will have a momentary visual change, but this change may return to normal in the next second. If we rely solely on visual judgment, a large number of false positive signals will be generated. This invention addresses the disconnect between visual images and operating parameters by integrating process conditions and physical dynamic constraints into visual decision-making in the A5 stage. First, in step A51, process condition features are extracted in parallel from two time series: one line is extracted from temperature time series data, using a gated loop unit to capture the dynamic trends and periodic characteristics of temperature changes; the other line is extracted from stirring operation data, using a one-dimensional convolutional network to identify the pattern characteristics of time series conditions such as stirring speed and stirring intensity—temperature affects the long-term evolution of reaction rates, and stirring affects the mixing and tumbling of the melt, requiring different time receptive fields to capture. The features extracted from the two lines are then fused, and a multilayer perceptron is used to generate the final process condition features, reflecting the comprehensive state of the melting process at the current moment in terms of both temperature and stirring dimensions. Subsequently, a process-visual affinity weight map is calculated. The affinity weight represents the proportion of visual features that are reliable signals of reaction progress under the current process conditions, and the proportion that may be false signals caused by changes in the conditions. This weight allows features at different locations in the image to receive differentiated confidence assessments. For example, under stable temperature and uniform stirring conditions, any visual color change will be given a high weight, as it is considered likely to be a genuine manifestation of the impurity removal reaction. However, during periods of drastic temperature or stirring fluctuations, even if a color change is observed, its weight will be reduced, because the change may simply be a disturbance of the conditions. By weighting the visual features using affinity weights, a condition-aligned visual feature is obtained. Since this feature has been calibrated, it is closer to the actual impurity removal reaction progress, rather than being misled by fluctuations in the conditions. In step A52, a spatial distribution matrix of reaction intensity is further established, embedding the constraints of the physical field into the visual feature space—the visual features aligned with the operating conditions are spatially fused with the process operating condition features, and a matrix reflecting the reaction intensity distribution at each location of the molten pool is constructed through convolution operations; this matrix is ​​no longer a purely visual feature, but a spatial representation of the reaction process under the dual constraints of vision and operating conditions; then, global average pooling is used to aggregate the spatial information of the matrix into a scalar, and the reaction progression is calculated through nonlinear mapping; the reaction progression is represented by a value between 0 and 1, reflecting the proportion of the impurity removal reaction from start to finish; In step A53, chemical reaction kinetics constraints are introduced. The critical activation level for burn-off is calculated using the reaction progress, the current molten pool temperature, and a physical model based on a modified Arrhenius equation. The Arrhenius equation describes the exponential change in chemical reaction rate with temperature—the higher the temperature, the faster the chemical reaction rate. This means that under high-temperature conditions, the oxidation reaction of bismuth metal will also accelerate, increasing the risk of burn-off. The calculation of the critical activation level for burn-off integrates the current reaction progress and temperature conditions, providing a physical assessment of whether bismuth metal has entered the burn-off risk zone at a specific process point. Finally, the critical activation level for burn-off is fused with the global characteristics of the reaction intensity spatial distribution matrix to output the critical risk value for excessive burn-off of bismuth metal. Compared to existing technologies, this invention establishes a process-visual affinity mechanism in the A51 stage, enabling dynamic alignment between visual features and process conditions—trusting vision under high-confidence conditions and reducing the influence weight of vision under low-confidence conditions, thus suppressing false positive signals. By embedding physical field constraints in the A52 stage, vision and process conditions are integrated into a comprehensive representation that includes both image information and process constraints. The A53 stage introduces chemical kinetic constraints, using the Arrhenius equation to directly encode the physical laws of temperature and reaction rate into the risk assessment. This allows the model to no longer judge burnout simply because it looks like burnout, but to comprehensively assess risk based on physical logic such as "at this temperature, at this stage of the reaction, what is the probability of a kinetic burnout reaction?" This avoids the limitations of pure AI vision in terms of its lack of understanding of physics, ultimately enabling more accurate and robust judgments in process decision-making.

[0017] Furthermore, step A6 includes: ; in, This serves as a marker for determining the termination of the reaction. To reflect the progress, This is the lower limit of the reaction extent, which terminates the reaction. This represents the critical risk value for excessive bismuth metal burn-off. These are the process prior parameters (burn-off risk safety thresholds). This is the lower limit of the safe temperature range.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention addresses the problem of information masking in the high-temperature alkaline visual field environment during lead-bismuth alloy refining. It constructs a multi-level vision-process fusion framework adapted to strong interference visual fields, effectively balancing the accuracy of impurity removal and the control of metal burn-off. First, through the decoupling and parallel stripping mechanism of molten pool interference, it specifically penetrates the smoke and dust masking, corrects thermal distortion, and suppresses thermal radiation color deviation, restoring the pure alkaline slag characteristics under complex working conditions, ensuring the fidelity of microscopic features such as the particle feel and looseness of the alkaline slag. On this basis, through multi-dimensional feature extraction and residual risk assessment, it achieves the separation of alkaline slag color gradient, viscosity, and burn-off characteristics, eliminating visual coupling interference between various process indicators. In addition, this invention dynamically aligns visual perception with real-time process parameters such as temperature and stirring, and introduces chemical reaction kinetic constraints. It uses physical laws to calibrate the confidence of visual signals, suppressing false positive signals caused by working condition fluctuations, and achieving objective quantification of the reaction progress and burn-off risk. It gets rid of the limitations of relying solely on experience and simple visual perception, providing reliable support for the robust control of the lead-bismuth alloy refining process.

[0019] (2) To address the problem that it is difficult to simultaneously and effectively eliminate the three types of composite interference with different physical properties—dust occlusion, thermal distortion, and color deviation—in a high-temperature alkaline smelting environment, this invention establishes an interference decoupling and stripping mechanism. By extracting features and calculating spatial gradients from the pre-processed alkaline slag image, the original pixel information is transformed into higher-order features, and then the dust occlusion probability map, thermal distortion probability map, and color deviation residual probability map are calculated to quantify the pollution degree and distribution location of various interferences in the image with pixel-level precision. Based on this, this invention designs three parallel interference stripping branches: the dust stripping branch gradually eliminates the dust-covered area through adaptive feature transformation and weighted suppression. The interference information is removed; the thermal distortion stripping branch corrects the geometric distortion caused by thermal convection and restores the outline and shape of the alkali slag; the color deviation stripping branch reweights the color information distorted by thermal radiation red light through the color channel and modulates the color information to highlight the true color of the alkali slag; compared with the limitations of traditional single global enhancement or general denoising algorithms that are difficult to deal with the three types of interference separately, the multi-channel parallel design of this invention ensures that each type of interference has a targeted elimination mechanism, avoiding the problems of incomplete interference elimination and over-processing that cause the fine particle texture and edge morphology information of the alkali slag to be smoothed or blurred, providing a more realistic and reliable visual input for subsequent alkali slag feature extraction and process decision.

[0020] (3) To address the problem that details such as the looseness of alkali residue accumulation, the thickness of the surface gas barrier layer, and connectivity are easily smoothed or blurred during the fusion of features from multiple branches in the three-channel parallel stripping process, leading to inaccurate process decisions, this invention designs an adaptive fusion mechanism. First, the interference coupling degree fusion weight tensor reflects the relative intensity relationship of the three types of interference at different locations, and spatially weighted fusion is performed on the output features of the three branches to ensure that each pixel position can obtain the most suitable interference elimination result. Second, a texture-preserving mask is introduced, and the gradient similarity between the interference stripping features and the original basic features is calculated to evaluate the edge and texture features of the alkali residue pixel by pixel. The invention generates an adaptive soft mask based on the retention status of the original features. This mask tends to trust the interference stripping results in regions with high gradient similarity, while retaining the details of the original features in regions with low similarity. Finally, the original features and stripped features are adaptively mixed through the mask to ensure that the feature image of the pure molten pool alkali slag is free from obvious interference and contamination, and that no key features are lost. Compared with traditional fixed-weight or simple weighted average fusion methods, this invention has spatial adaptability and can perform differentiated processing according to the interference intensity distribution in different regions. It can directly detect whether the key information of the alkali slag itself is retained, providing a high-fidelity visual basis for judging the degree of impurity removal and controlling the process sequence.

[0021] (4) In view of the mutual influence of the characteristics of the three process judgment dimensions of alkaline slag color gradient, viscosity, and bismuth oxidation burn-off, which makes it difficult to accurately distinguish whether the color change is due to impurity removal or burn-off, this invention adopts a multi-dimensional collaborative feature extraction and risk assessment mechanism. Each of the three process judgment dimensions is set up with an independent feature extraction branch: the color gradient feature extraction branch focuses on capturing the color distribution and gradient trend of alkaline slag; the viscosity texture feature extraction branch focuses on capturing visual indicators related to viscosity, such as the particle feel and fluidity of alkaline slag; the bismuth oxidation burn-off feature extraction branch focuses on capturing local pixel-level information related to metal oxidation characteristics; at the same time, by calculating the interference residual risk coefficient map, the threat of residual interference at each location to the credibility of the feature is assessed, and the feature is weighted and weakened to prevent confusing information from misleading subsequent decisions; finally, through collaborative gating weighting, the representativeness of the current fused feature to the real alkaline slag reaction state is reflected, realizing the multi-dimensional decoupling extraction of features, and through secondary interference filtering, the model can more accurately understand the real process state of the molten pool.

[0022] (5) To address the problem of false positive interference signals caused by fluctuations in operating conditions during high-temperature smelting, this invention integrates process conditions and physical dynamic constraints into visual judgment. First, process condition features are extracted in parallel from two time series: temperature and stirring. The dynamic trend of temperature change and the pattern features of stirring are captured respectively, and the fusion reflects the comprehensive state of the smelting process. Then, the process-visual affinity weight map is calculated. Under high-confidence conditions with stable temperature and uniform stirring, visual features are trusted, while under low-confidence conditions with drastic fluctuations, the weight of visual influence is reduced, thereby achieving dynamic matching between visual features and process conditions. Furthermore, a reaction intensity spatial distribution matrix is ​​established, and physical field constraints are embedded into the visual feature space to obtain a comprehensive characterization that includes both image information and process constraints. Finally, chemical reaction kinetic constraints are introduced, and based on a physical model derived from the Arrhenius formula, the critical activation amount for burn-off is calculated by comprehensively considering the degree of reaction progression and temperature conditions, thereby outputting the critical risk value for excessive burn-off of bismuth metal. Compared with the limitations of existing technologies that rely solely on a single visual feature, this invention suppresses false positive signals through a process-visual affinity mechanism and avoids decision bias due to a lack of understanding of physics through physical and kinetic constraints, making process decisions more accurate and robust. Attached Figure Description

[0023] Figure 1 A flowchart illustrating an intelligent control method for the smelting and refining process of a lead-bismuth alloy provided by the present invention; Figure 2 Interference residual risk coefficient diagram provided for this invention. Detailed Implementation

[0024] The present invention will be further described below with reference to the accompanying drawings, but this is not intended to limit the present invention in any way. Any modifications or substitutions made based on the teachings of the present invention shall fall within the protection scope of the present invention.

[0025] Example 1: A smart control method for the smelting and refining process of lead-bismuth alloy, such as... Figure 1 As shown, it includes the following steps: A1: Collect visual images of the alkaline slag, melting temperature data, and stirring operation data of the lead-bismuth alloy alkaline refining molten pool, and preprocess them respectively to obtain preprocessed visual images of the alkaline slag, preprocessed melting time-series temperature data, and preprocessed stirring operation time-series data, including: A11: Visual images of alkaline slag in the alkaline refining pool of lead-bismuth alloy were acquired by an industrial high-temperature dustproof camera device. Then, Gaussian filtering for noise reduction, optical distortion correction, and thermal radiation color shift compensation were applied sequentially to obtain the pre-processed visual images of alkaline slag. A12: The melting temperature data of the lead-bismuth alloy refining furnace is collected by armored thermocouples. The data type is one-dimensional time-series numerical data. The melting temperature data of the melting pool is processed by removing three times the standard difference constant and completing the missing linear interpolation to obtain the preliminary pre-processed melting time-series temperature data of the melting pool. A13: The stirring operation data of the refining furnace stirring actuator is collected through the stirring mechanism condition acquisition module. The data type is one-dimensional time-series numerical data, including the real-time stirring speed, stirring start and stop status, and stirring operation sequence cycle. The stirring operation data is processed by three times standard difference constant value removal and time sequence normalization to obtain the preliminary pre-processed stirring operation sequence data. A14: The visual images of the pre-treated alkali slag, the molten pool smelting time-series temperature data, and the stirring operation sequence data after the preliminary pretreatment are synchronized with timestamps. Linear interpolation or spline interpolation methods are used to resample all data to a uniform sampling frequency to obtain the visual images of the pre-treated alkali slag, the molten pool smelting time-series temperature data, and the stirring operation sequence data after the preliminary pretreatment.

[0026] A2: Based on the pre-processed visual image of the alkali slag, the shallow general features of the furnace image, the probability map of smoke and dust obstruction, the probability map of thermal distortion, and the probability map of color deviation residue are extracted sequentially. Then, through a three-channel parallel stripping mechanism to decouple the molten pool interference, the output features of the smoke and dust stripping branch, the output features of the thermal distortion stripping branch, and the output features of the color deviation stripping branch are extracted, including: A21: Based on the preprocessed visual image of the alkali slag, shallow general features of the furnace image are extracted using a convolutional encoder with residual connections. The probability maps of smoke occlusion, thermal distortion, and color cast residue at each pixel location are calculated using point-by-point convolution and channel segmentation. The calculation method is as follows: ; ; ; ; ; in, This is a common feature of shallow layers in the furnace surface. For convolutional layers with residual connections, Visual images of the pretreated alkali residue. The spatial gradient features are general shallow features of the furnace image. For the Sobel edge gradient operator, The fusion features after splicing This is for channel splicing operations. To fuse the interference probability tensor, For the Softmax function, For pointwise convolutional layers, This is an operation to divide the channel equally along its dimension. A probability diagram showing the obstruction caused by smoke and dust. This is a thermal distortion probability map. This is a probability diagram of color deviation residue. A22: Based on the shallow general features of the furnace image and the probability maps of smoke and dust obstruction, thermal distortion, and color deviation residue, a three-channel parallel stripping mechanism for decoupling molten pool interference is used to strip away smoke and dust obstruction, thermal convection distortion, and color deviation residue respectively, resulting in the output features of the smoke and dust stripping branch, the thermal distortion stripping branch, and the color deviation stripping branch. The calculation method is as follows: ; ; ; in, The output characteristics of the dust stripping branch are as follows. For element-wise subtraction, For dimensional expansion operations, For Hadama accumulation, For dilated convolution operators, This represents the output characteristics of the thermal distortion stripping branch. It is a convolutional layer. The output features of the color-bias stripping branch are as follows: This is a channel attention mechanism.

[0027] A3: Based on the output characteristics of the dust stripping branch, the thermal distortion stripping branch, the color shift stripping branch, as well as the dust occlusion probability map, thermal distortion probability map, color shift residue probability map, and shallow general features of the furnace image, adaptive fusion is performed to sequentially calculate the preliminary interference stripping features, the alkaline slag texture protection soft mask, and the pure molten pool alkaline slag feature image, including: A31: Based on the output characteristics of the smoke and dust stripping branch, the output characteristics of the thermal distortion stripping branch, the output characteristics of the color shift stripping branch, the smoke and dust occlusion probability map, the thermal distortion probability map, and the color shift residue probability map, the three stripping results are adaptively fused to obtain preliminary interference stripping characteristics. The calculation method is as follows: ; ; ; in, To fuse the weight tensor for interference coupling, This is a convolutional layer with a kernel size of 3×3. Spatial fusion weight map of smoke and dust stripping branches. This is a spatial fusion weight map of the thermal distortion stripping branch. For the spatial fusion weight map of the color bias stripping branch, This is an operation to divide the channel equally along its dimension. Preliminary interference stripping characteristics, This is element-wise addition; A32: Based on the initial interference stripping features and the shallow general features of the furnace image, an alkaline slag texture protection soft mask is calculated using an alkaline slag texture fidelity adaptive mask refinement mechanism. The initial interference stripping features are then refined to obtain a pure molten pool alkaline slag feature image. The calculation method is as follows: ; ; ; ; in, The spatial gradient features are the initial interference stripping features. To standardize gradient similarity, It is a very small positive number. A soft mask is used to protect the texture of alkali residue. For the Sigmoid function, This is the gradient similarity amplification factor. This is a characteristic image of alkaline slag from a pure molten pool.

[0028] A4: Based on the characteristic image of the alkaline slag from the pure molten pool, extract the color gradient features, thinning / condensation texture features, and bismuth oxidation burn-off features of the alkaline slag in sequence. Calculate the interference residue risk coefficient map, the visual features of the fused alkaline slag, and the visual features of the effective alkaline slag, including: A41: Based on the characteristic image of alkaline slag from a pure molten pool, the color gradient features, thinning / dense texture features, and bismuth oxidation burn-off features of the alkaline slag are extracted using a multi-scale collaborative gated feature fusion mechanism. The calculation method is as follows: ; ; ; in, The color of the alkali residue is characterized by a gradual change. For residual blocks with residual connections, For hollow convolution, It exhibits the characteristic thin-to-thick texture of alkali residue. Characterized by bismuth oxidation and burn-off; A42: Based on the color gradient characteristics, viscosity texture characteristics, and bismuth oxidation burn-off characteristics of the alkali slag, the interference residue risk coefficient at each spatial location is assessed, resulting in an interference residue risk coefficient map. The visual characteristics of the fused alkali slag are then calculated using the following method: ; ; ; ; in, For the residual risk coefficient diagram of interference, These are the color component, texture component, and burn-off component, respectively. This is a feature that mitigates the risk of interference. Visual characteristics of the fused alkali residue; interference residue risk coefficient diagram as shown in the figure. Figure 2 As shown; among them, the discretely distributed bright patches represent the risk of local interference caused by thermal wave distortion in the center of the molten pool, while the dark low-value areas occupying most of the image area represent clear observation windows where the visual characteristics of the alkali slag are effectively preserved. A43: Based on the fused visual characteristics of the alkali residue, weighted fusion is performed through collaborative gating to obtain the effective visual characteristics of the alkali residue. The calculation method is as follows: ; ; in, To coordinate gating weights, It is a multilayer perceptron. For global average pooling, The visual characteristics of the fused alkali residue To effectively identify the visual characteristics of alkali residue, This is a residual block with residual connections.

[0029] A5: Based on the effective visual characteristics of the alkali slag, the smelting time-series temperature data of the pretreated molten pool, and the stirring operation time-series data of the pretreated pool, calculate the reaction advancement degree and the critical risk value of excessive bismuth metal burn-off, including: A51: For each frame of preprocessed alkali residue visual image, the effective alkali residue visual features are determined based on their timestamps. The preprocessed molten pool melting time-series temperature data and the preprocessed stirring time-series operating condition data were used to extract process condition features using gated recurrent units and one-dimensional convolutional networks, respectively. Then, the process-visual affinity weight map and the visual features aligned with the operating conditions were calculated. The calculation method is as follows: ; ; ; in, For process operating conditions, For linear layers, For gated loop unit, This is the pre-processed molten pool melting time-series temperature data. It is a one-dimensional convolutional network. This is the pre-processed stirring operation sequence data. For the process-visual affinity weighting diagram, For outer product operation, Align visual features with operating conditions; A52: Based on the visual features aligned with the operating conditions and the characteristics of the process conditions, a spatial distribution matrix of the reaction intensity is established. The dynamic constraints of the physical field are embedded into the visual feature space, and the reaction advancement degree is calculated using global average pooling and nonlinear mapping kernels. The calculation method is as follows: ; ; in, The reaction intensity spatial distribution matrix, For ReLU function, To reflect the degree of advancement; A53: Based on the reaction progression, the pre-treated molten pool melting time-series temperature data, and the reaction intensity spatial distribution matrix, the critical activation amount for burn-off is calculated, and then the critical risk value for excessive bismuth metal burn-off is assessed and output. The calculation method is as follows: ; ; in, This is the critical activation threshold for burn-off. The material burn-off coefficient, It is an exponential function. For activation energy constraint parameters, To retrieve the last value, This represents the critical risk value for excessive bismuth metal burn-off.

[0030] A6: Based on the reaction progress, the critical risk value of excessive bismuth metal burn-off, and the smelting time-series temperature data of the pretreated molten pool, assess the termination conditions of the refining reaction, obtain the reaction termination judgment flag, and terminate the smelting reaction when the reaction termination judgment flag is 1, including: ; in, This serves as a marker for determining the termination of the reaction. This is the lower limit of the reaction extent, which terminates the reaction. This represents the critical risk value for excessive bismuth metal burn-off. These are the process prior parameters (burn-off risk safety thresholds). This is the lower limit of the safe temperature range.

[0031] Specifically, for scenarios where molten pool thermal convection causes severe fluctuations in localized reactions, this invention also provides an adaptive termination determination method based on risk-progress dynamics balance to replace step A6. The calculation method is as follows: ; in, As a risk inhibitor, This is the threshold for process coordination termination.

[0032] In the implementation of this invention, the specific parameter settings and training configurations of each neural network module are as follows: The convolutional encoder and probabilistic graph computation in step A2: The convolutional encoder with residual connections uses three stacked convolutional blocks, with 64, 128, and 256 neurons in each layer, respectively. The kernel size is 3×3, the stride is 1, the padding uses the same mode, and the activation function is ReLU. The Sobel edge gradient operator uses a fixed Sobel convolutional kernel for gradient calculation. The number of output channels of the pointwise convolutional layer in the probabilistic graph fusion process is set to 3. The three-channel parallel stripping mechanism in step A2: The dilated convolution in the smoke and dust stripping branch uses standard convolution with a kernel size of 3×3 and 128 neurons, followed by a dilated convolution with a dilation rate of 2 and a kernel size of 3×3, also with 128 neurons; the convolutional layer in the thermal distortion stripping branch uses a kernel size of 3×3 and 256 neurons; the channel attention mechanism in the color deviation stripping branch uses global average pooling followed by two fully connected layers. The first fully connected layer has 64 neurons and uses ReLU as the activation function, while the second fully connected layer has the same number of neurons as the input channels and uses Sigmoid as the activation function. Step A3: The convolutional layer for interference coupling fusion uses a kernel size of 3×3, 256 neurons, and 3 output channels to segment the three types of interference weights; the fine convolutional layer in the alkaline slag texture protection soft mask uses a kernel size of 1×1, 128 neurons, a gradient similarity amplification factor α of 4.0, and a minimum normal number e_min of 1e-8. Step A4: The color gradient feature extraction branch uses a combination of standard convolutional layers and residual blocks. The convolutional layers have a kernel size of 3×3 and 128 neurons. The residual blocks contain two convolutional layers, each with 128 neurons, and the activation function is ReLU. The sparse and dense texture feature extraction branch uses dilated convolutions with a dilation rate of 3, a kernel size of 3×3, and 128 neurons. The bismuth oxidation burn-off feature extraction branch uses pointwise convolutions with a kernel size of 1×1 and 128 neurons. The convolutional layer for interference residual risk assessment has a kernel size of 1×1, 256 neurons, and 1 output channel. The adaptive gated convolutional layer in the fusion process has a kernel size of 3×3 and 256 neurons. The multilayer perceptron in the collaborative gating contains two fully connected layers. The first layer has 128 neurons and the activation function is ReLU, and the second layer has 1 neuron and the activation function is Sigmoid. Step A5: The number of hidden layer neurons in the gated recurrent unit (GRU) is set to 128. The one-dimensional convolutional network uses 4 layers of convolution with kernel sizes of 5, 7, 11, and 15, each with 128 neurons and a stride of 1. The padding uses the same mode. The multilayer perceptron for process condition fusion contains two fully connected layers. The first layer has 128 neurons and uses ReLU as the activation function. The second layer has 256 neurons. In the calculation of the spatial distribution matrix of reaction intensity, two layers of dilated convolution with an inflation rate of 2 and a kernel size of 3×3 are used, with 256 neurons. The receptive field is expanded to 7×7 to fully capture the global reaction distribution while avoiding too many parameters. The multilayer perceptron for calculating the reaction progression degree and the critical activation amount of burnout adopts a unified design: the first fully connected layer outputs 256 neurons with ReLU as the activation function; the second fully connected layer outputs 128 neurons with ReLU as the activation function; finally, a linear layer outputs a scalar value, followed by Sigmoid activation to obtain a probability value in the range [0,1].

[0033] The entire neural network architecture is trained using an end-to-end supervised learning approach. The training data consists of multimodal data collected during the actual lead-bismuth alloy refining process, including images of alkaline slag labeled with reaction termination and burn-off risk tags, temperature time-series data, and stirring condition data. The optimizer uses the Adam optimizer with an initial learning rate of 1e-4, employing a cosine annealing decay strategy, and a total of 100 training epochs. To adapt to the characteristics of industrial data, the batch size is set to 16, and a gradient accumulation strategy is used to update weights every 4 batches. The loss function uses a task uncertainty weighting method to remove interference. In addition to the loss (smoke removal, thermal distortion correction, color shift modulation), the loss for predicting the degree of reaction progression and the loss for assessing burn risk, weight parameters were learned for each of these losses. The initial weights were 1.0, 1.5, and 2.0, respectively. During training, the weight parameters were optimized by minimizing the total weighted loss. Data augmentation techniques were used during training to transform the alkaline residue image by random rotation, scaling, and color temperature changes to enhance the robustness and generalization ability of the model. The validation set consisted of 10% of the training data, and an early stopping mechanism was used to monitor the validation loss. Training was stopped when the validation loss decreased by less than 1e-4 within 15 consecutive rounds.

[0034] It should be noted that the sequence numbers of the above embodiments of the present invention are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, apparatus, article, or method. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0035] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0036] The above are merely preferred embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for intelligent control of the lead-bismuth alloy smelting and refining process, characterized in that, Includes the following steps: A1: Collect visual images of alkaline slag, melting temperature data, and stirring operation data of the alkaline refining pool of lead-bismuth alloy, and preprocess them respectively to obtain preprocessed visual images of alkaline slag, preprocessed melting time-series temperature data, and preprocessed stirring operation sequence data. A2: Based on the pre-processed visual image of the alkali slag, the shallow general features of the furnace image, the probability map of smoke and dust obstruction, the probability map of thermal distortion, and the probability map of color deviation residue are extracted in sequence. Then, through the three-channel parallel stripping mechanism of molten pool interference decoupling, the output features of the smoke and dust stripping branch, the output features of the thermal distortion stripping branch, and the output features of the color deviation stripping branch are extracted. A3: Based on the output characteristics of the dust stripping branch, the thermal distortion stripping branch, the color deviation stripping branch, as well as the dust occlusion probability map, the thermal distortion probability map, the color deviation residual probability map, and the shallow general features of the furnace image, adaptive fusion is performed to calculate the preliminary interference stripping features, the alkaline slag texture protection soft mask, and the pure molten pool alkaline slag feature image in sequence. A4: Based on the characteristic image of alkaline slag in the pure molten pool, extract the color gradient feature, thin-thickness texture feature, and bismuth oxidation burn-off feature of the alkaline slag in sequence, and calculate the interference residual risk coefficient map, the visual features of the fused alkaline slag, and the visual features of the effective alkaline slag. A5: Based on the effective visual characteristics of alkali slag, the molten pool smelting time-series temperature data after pretreatment, and the stirring operation sequence data after pretreatment, calculate the reaction advancement degree and the critical risk value of excessive bismuth metal burn-off. A6: Based on the reaction progress, the critical risk value of excessive bismuth metal burn-off, and the smelting time sequence temperature data of the pretreated molten pool, evaluate the termination conditions of the refining reaction, obtain the reaction termination judgment mark, and terminate the smelting reaction when the reaction termination judgment mark is 1.

2. The intelligent control method for the lead-bismuth alloy smelting and refining process according to claim 1, characterized in that, Step A1 includes: A11: Visual images of alkaline slag in the alkaline refining pool of lead-bismuth alloy were acquired by an industrial high-temperature dustproof camera device. Then, Gaussian filtering for noise reduction, optical distortion correction, and thermal radiation color shift compensation were applied sequentially to obtain the pre-processed visual images of alkaline slag. A12: The melting temperature data of the lead-bismuth alloy refining furnace is collected by armored thermocouples. The data type is one-dimensional time-series numerical data. The melting temperature data of the melting pool is processed by removing three times the standard difference constant and completing the missing linear interpolation to obtain the preliminary pre-processed melting time-series temperature data of the melting pool. A13: The stirring operation data of the refining furnace stirring actuator is collected through the stirring mechanism condition acquisition module. The data type is one-dimensional time-series numerical data, including the real-time stirring speed, stirring start and stop status, and stirring operation sequence cycle. The stirring operation data is processed by three times standard difference constant value removal and time sequence normalization to obtain the preliminary pre-processed stirring operation sequence data. A14: The visual images of the pre-treated alkali slag, the molten pool smelting time-series temperature data, and the stirring operation sequence data after the preliminary pretreatment are synchronized with timestamps. Linear interpolation or spline interpolation methods are used to resample all data to a uniform sampling frequency to obtain the visual images of the pre-treated alkali slag, the molten pool smelting time-series temperature data, and the stirring operation sequence data after the preliminary pretreatment.

3. The intelligent control method for the lead-bismuth alloy smelting and refining process according to claim 1, characterized in that, Step A2 includes: A21: Based on the preprocessed visual image of the alkali slag, shallow general features of the furnace image are extracted using a convolutional encoder with residual connections. The probability maps of smoke occlusion, thermal distortion, and color cast residue at each pixel location are calculated using point-by-point convolution and channel segmentation. The calculation method is as follows: ； ； ； ； ； in, This is a common feature of shallow layers in the furnace surface. For convolutional layers with residual connections, Visual images of the pretreated alkali residue. The spatial gradient features are general shallow features of the furnace image. For the Sobel edge gradient operator, The fusion features after splicing This is for channel splicing operations. To fuse the interference probability tensor, For the Softmax function, For pointwise convolutional layers, This is an operation to divide the channel equally along its dimension. A probability diagram showing the obstruction caused by smoke and dust. This is a thermal distortion probability map. This is a probability diagram of color deviation residue. A22: Based on the shallow general features of the furnace image and the probability maps of smoke and dust obstruction, thermal distortion, and color deviation residue, a three-channel parallel stripping mechanism for decoupling molten pool interference is used to strip away smoke and dust obstruction, thermal convection distortion, and color deviation residue respectively, resulting in the output features of the smoke and dust stripping branch, the thermal distortion stripping branch, and the color deviation stripping branch. The calculation method is as follows: ； ； ； in, The output characteristics of the dust stripping branch are as follows. For element-wise subtraction, For dimensional expansion operations, For Hadama accumulation, For dilated convolution operators, This represents the output characteristics of the thermal distortion stripping branch. It is a convolutional layer. The output features of the color-bias stripping branch are as follows: This is a channel attention mechanism.

4. The intelligent control method for the lead-bismuth alloy smelting and refining process according to claim 3, characterized in that, Step A3 includes: A31: Based on the output characteristics of the smoke and dust stripping branch, the output characteristics of the thermal distortion stripping branch, the output characteristics of the color shift stripping branch, the smoke and dust occlusion probability map, the thermal distortion probability map, and the color shift residue probability map, the three stripping results are adaptively fused to obtain preliminary interference stripping characteristics. The calculation method is as follows: ； ； ； in, To fuse the weight tensor for interference coupling, This is a convolutional layer with a kernel size of 3×3. Spatial fusion weight map of smoke and dust stripping branches. This is a spatial fusion weight map of the thermal distortion stripping branch. For the spatial fusion weight map of the color bias stripping branch, This is an operation to divide the channel equally along its dimension. Preliminary interference stripping characteristics, This is element-wise addition; A32: Based on the initial interference stripping features and the shallow general features of the furnace image, an alkaline slag texture protection soft mask is calculated using an alkaline slag texture fidelity adaptive mask refinement mechanism. The initial interference stripping features are then refined to obtain a pure molten pool alkaline slag feature image. The calculation method is as follows: ； ； ； in, The spatial gradient features are the initial interference stripping features. To standardize gradient similarity, It is a very small positive number. A soft mask is used to protect the texture of alkali residue. For the Sigmoid function, This is the gradient similarity amplification factor. This is a characteristic image of alkaline slag from a pure molten pool.

5. The intelligent control method for the lead-bismuth alloy smelting and refining process according to claim 4, characterized in that, The A4 step includes: A41: Based on the characteristic image of alkaline slag from a pure molten pool, the color gradient features, thinning / dense texture features, and bismuth oxidation burn-off features of the alkaline slag are extracted using a multi-scale collaborative gated feature fusion mechanism. The calculation method is as follows: ； ； ； in, The color of the alkali residue is characterized by a gradual change. For residual blocks with residual connections, For hollow convolution, It exhibits the characteristic thin-to-thick texture of alkali residue. Characterized by bismuth oxidation and burn-off; A42: Based on the color gradient characteristics, viscosity texture characteristics, and bismuth oxidation burn-off characteristics of the alkali slag, the interference residue risk coefficient at each spatial location is assessed, resulting in an interference residue risk coefficient map. The visual characteristics of the fused alkali slag are then calculated using the following method: ； ； ； ； in, For the residual risk coefficient diagram of interference, These are the color component, texture component, and burn-off component, respectively. This is a feature that mitigates the risk of interference. Visual characteristics of the fused alkali residue; A43: Based on the fused visual characteristics of the alkali residue, weighted fusion is performed through collaborative gating to obtain the effective visual characteristics of the alkali residue. The calculation method is as follows: ； ； in, To coordinate gating weights, It is a multilayer perceptron. For global average pooling, The visual characteristics of the fused alkali residue To effectively identify the visual characteristics of alkali residue, This is a residual block with residual connections.

6. The intelligent control method for the lead-bismuth alloy smelting and refining process according to claim 5, characterized in that, Step A5 includes: First, the pre-processed molten pool melting time-series temperature data is processed by a gated loop unit and then by a linear layer; the pre-processed stirring running time-series condition data is processed by a one-dimensional convolutional network and then by a linear layer; the above two processing results are spliced ​​together and then processed by a multilayer perceptron to obtain the process condition characteristics. The effective alkali residue visual features are processed by a convolutional layer with a kernel size of 1×1, and then the outer product operation is performed with the process condition features. After processing by the Sigmoid function, the process-visual affinity weight map is obtained. The process-visual affinity weight map is then multiplied by the Hadamard product with the effective alkali residue visual features to obtain the condition-aligned visual features. The visual features aligned with the working conditions are added element by element to the expanded process condition features, and then processed sequentially by a convolutional layer with a kernel size of 3×3 and a ReLU function to obtain the spatial distribution matrix of the reaction intensity. The spatial distribution matrix of the reaction intensity is then processed by global average pooling, and then processed sequentially by a multilayer perceptron and a Sigmoid function to obtain the degree of reaction advancement. The critical activation level for burn-off is obtained by subtracting the product of the material burn-off coefficient and the exponential function from the reaction advancement rate, followed by processing with the Sigmoid function; where the power of the exponential function is: ； in, For activation energy constraint parameters, To retrieve the last value, This refers to the pre-processed molten pool smelting timing temperature data; The critical activation value for bismuth burn-off is obtained by concatenating the results of global average pooling processing with the spatial distribution matrix of reaction intensity, followed by processing by multilayer perceptron and sigmoid function.

7. The intelligent control method for the lead-bismuth alloy smelting and refining process according to claim 6, characterized in that, Step A6 includes: ； in, This serves as a marker for determining the termination of the reaction. To reflect the progress, This is the lower limit of the reaction extent, which terminates the reaction. This represents the critical risk value for excessive bismuth metal burn-off. These are the process prior parameters (burn-off risk safety thresholds). This is the lower limit of the safe temperature range.

Images