A visual detection method for water outlet of silicon water of an electric arc furnace

Abstract

The application discloses a visual detection method for silicon water of a submerged arc furnace, which comprises the following steps: a double-flow feature extraction step, in which appearance and motion deep features are extracted from continuous video frame images of the submerged arc furnace, and multi-scale detail feature enhancement is performed; a space-time feature modulation step, in which spatial attention weight is calculated through motion feature, and appearance features are modulated to form fusion features; a fluid feature constraint step, in which a gravity coherence loss module is introduced into a feature space, a vertical flow gradient penalty is calculated, and corresponding physical coherence constraints are generated; a network joint optimization step, in which a lightweight encoder-decoder is used to optimize network loss of the fusion features; and a silicon water state detection step, in which the optimized features are input into a segmentation network, and a silicon water outflow state is positioned and discriminated through connected domain analysis and morphological filtering. The application can effectively identify continuous flow features of silicon water of the submerged arc furnace under the working conditions of light change and smoke shielding, reduce background interference, and improve detection precision and efficiency.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and metallurgical automation monitoring technology, and more specifically to a visual detection method for silicon molten water from a submerged arc furnace. Background Technology

[0002] As a core smelting equipment in the industrial silicon and ferroalloy metallurgical industries, the status monitoring of the furnace outlet (silicon molten metal outflow) is crucial for production process control, output assessment, and on-site safety management. Its automation level and equipment operating efficiency directly depend on the accuracy and real-time performance of the outflow status identification. When using machine vision-based image detection methods to detect the outflow status of submerged arc furnaces, it is essential to ensure that the fluid features in the acquired images or videos are complete and clear. However, in actual high-temperature smelting industrial environments, due to factors such as extremely strong glare, high-temperature thermal radiation interference, large amounts of dense smoke and dust obscuring the image, and complex fluctuations in workshop lighting, the acquired furnace outlet video images often suffer from severe local overexposure and blurred fluid edges. This significantly affects the stability and reliability of outflow status monitoring.

[0003] Currently, the main methods for detecting the status of molten silicon outlet water in submerged arc furnaces include manual on-site observation and techniques based on traditional image processing. Manual on-site observation not only involves harsh working environments and poses significant safety hazards, but is also inefficient, highly subjective, and prone to misjudgments or missed detections, failing to meet the demands of modern metallurgical industries for efficient automated monitoring. Traditional image processing methods, such as simple threshold segmentation and background subtraction, have limited adaptability to complex high-temperature scenarios with static luminescent slag or dynamic obstruction by dense smoke, and their ability to identify continuous fluids and splashing sparks is also insufficient. Furthermore, while deep learning-based video detection methods have shown excellent accuracy in recent years, the extremely complex operating conditions of submerged arc furnaces and the variable flow patterns of molten silicon mean that these methods lack constraints on the physical continuity of the fluid in practical applications, easily leading to gaps or false alarms. Simultaneously, the large model structure results in extremely high computational requirements, making it difficult to achieve high frame rate real-time detection on edge computing devices in industrial settings with limited computing power. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide a visual detection method for molten silica water effluent from submerged arc furnaces. By introducing a dual-flow spatiotemporal feature modulation mechanism, combined with gravity coherence physical constraints and a lightweight network design strategy, this method solves the problems of misjudgment and fluid discontinuity that traditional methods are prone to under complex working conditions such as strong light reflection and dense smoke obstruction. At the same time, it significantly reduces the model's dependence on computing power and can strictly meet the production cycle requirements on edge devices in industrial sites with limited computing power, accurately locating and identifying the continuous flow status of molten silica water from submerged arc furnaces.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A visual inspection method for molten silica outlet water from a submerged arc furnace includes the following steps: The dual-stream feature extraction step involves acquiring continuous video frame images of the submerged arc furnace, extracting deep appearance features and deep motion features separately using a dual-stream encoder, and enhancing detail features through a multi-scale module. The spatiotemporal feature modulation step involves calculating spatial weights for the extracted motion deep features through an attention mechanism, and then dynamically modulating the appearance deep features to form fused features. The fluid feature constraint step introduces a gravity coherence loss module into the feature space to calculate the gradient penalty in the vertical flow direction and generate the corresponding physical coherence constraint. The network joint optimization step involves reconstructing the fused features using a lightweight codec, and optimizing the network by combining gravity coherence loss and multi-scale loss functions to enhance the physical consistency of fluid features. The silicon water state detection step involves inputting the optimized features into a segmentation network, and using connected component analysis and morphological filtering to locate and determine the outflow state of the silicon water.

[0006] Furthermore, the dual-stream encoder in the dual-stream feature extraction step is constructed based on a lightweight convolutional neural network architecture pre-trained on the ImageNet dataset, and its output feature map undergoes feature dimension compression and projection transformation through a lightweight channel adapter.

[0007] Furthermore, the spatiotemporal feature modulation step includes a motion-gated modulation strategy. This strategy involves generating spatial attention weights by activating deep motion features, and then multiplying these weights pixel-by-pixel with the deep appearance features to construct a fused spatiotemporal feature representation, the expression of which is: ; in, The output is the fused feature of the input image in the i-th frame. This is an element-wise multiplication operation between feature map channels. For activation function, For spatial feature convolution operation, For deep appearance features extracted from a dual-stream encoder, These are the deep motion features extracted from the dual-stream encoder.

[0008] Furthermore, the fluid characteristic constraint step also includes a gravity coherence loss function, which is used to calculate the characteristic distribution difference in the vertical flow direction to apply continuity constraints. Its expression is: ; in, To calculate the gravity coherence loss value, The preset gradient penalty function, Let (h, w) be the predicted output value of the feature map, and (h, w) be the spatial location. The predicted output value is the feature value of the position adjacent to (h,w) along the opposite direction of gravity.

[0009] Furthermore, the network joint optimization step includes a global reconstruction loss strategy and a local coherence loss strategy. The global reconstruction loss strategy includes measuring large-size fluid characteristics by calculating the difference in the basic segmentation distribution, and its expression is: ; The local coherence loss strategy controls the effective area of ​​gravity penalty through dynamic region mask constraints, and its expression is as follows: ; in, For the overall reconstruction loss, Let P be the total loss for local coherence, P be the network's predicted output features, and Y be the true state label. For adaptive constraint binary mask, The gravitational constraint strength coefficient. and These are the dynamic weighting coefficients corresponding to the global loss. This is due to the loss of local gravity continuity.

[0010] Furthermore, the network lightweight optimization step includes a lightweight codec construction strategy, which comprises a backbone decoding path and a cross-layer skip connection path. The backbone decoding path consists of a feature reconstruction module composed of lightweight convolutional layers, batch normalization, and activation functions. The cross-layer skip connection path introduces a high-resolution feature concatenation layer. When optimizing the lightweight codec network, backpropagation is performed by minimizing the joint expected loss function on the training dataset, the expression of which is: ; Where L is the final optimized joint expectation loss function, and (X,Y) is the loss function derived from the training dataset. Mid-sampled input video frame image pairs and their corresponding real-world mask labels. This represents the total loss due to local continuity issues.

[0011] Furthermore, the network joint optimization step also includes a dynamic feature discrimination and adversarial learning strategy. The adversarial learning strategy includes constructing a gated discrimination module. This gated discrimination module updates network weights based on a hinged discrimination loss function with boundary margin to distinguish between dynamic fluid features and static interference features. Its expression is: ; in, Here, is the dynamic feature discrimination loss function, where The probability distribution representing the dynamic characteristics of silicon water. The probability distribution representing the static highlighted background features. It is the output function of the gating discrimination module. The set boundary margin; through the adversarial learning strategy, the network is forced to push away the distribution boundary between the flowing silicon water and the static background in the feature space.

[0012] Furthermore, the silicon water state detection step includes morphological and geometric constraint strategies, which include independent connected component extraction of the initial feature mask, expressed as: ; Its arbitrary connected components Its features include an analytically computable geometric distribution: ; Where R is the set of extracted connected components, For connected component extraction operation, This serves as the initial binarization mask for the segmentation network output, where N is the total number of connected components. Let be the pixel area of ​​the k-th connected component. For spatial coordinates, The height of the bounding rectangle of the k-th connected region is given; then, the final silicon water positioning result is generated by combining the area and aspect ratio of the gravity-induced sagging shape with a joint threshold. ; in, To position the silica water as a mask after filtering out static noise. The preset minimum effective area threshold, Let the width of the bounding rectangle of the k-th connected component be . To prevent extremely small constants with a denominator of zero, This is the preset aspect ratio threshold.

[0013] Furthermore, the silica water state detection step also includes an inference and water outflow confidence assessment strategy. This strategy involves upsampling and interpolating the network output to the original video frame size and then smoothing it using Gaussian smoothing to obtain a pixel-level water outflow confidence score for precise silica water localization. The calculation process is as follows: ; in, Pixel-level water discharge confidence score. The initial probability score is the output of the segmentation network. and The Gaussian smoothing function and the upsampling interpolation function, respectively, are used to determine the frame-level water discharge state score for triggering system alarms. The maximum response value of all pixels in the array is given, and its expression is: ; in, The water discharge status score is at the frame level. Spatial coordinates The corresponding pixel-level output.

[0014] The beneficial effects of this invention are as follows: By introducing a dual-stream spatiotemporal feature modulation mechanism combined with gravity coherence physical constraints, it effectively solves the problem of fluid state detection under complex conditions such as strong light reflection and dense smoke obstruction. Specifically, dual-stream feature extraction and multi-scale enhancement strategies enrich the expression of motion and appearance features; a lightweight encoder-decoder combined with local gravity coherence loss ensures the integrity of the fluid physical structure and effectively avoids feature discontinuity; dynamic feature discrimination and adversarial learning mechanisms significantly improve the model's sensitivity to distinguishing real silicon-water boundaries in extremely bright regions. This method significantly reduces the dependence on edge device hardware computing power while ensuring extremely high detection accuracy. It not only has strong industrial-grade generalization ability but also meets the processing needs of real-time on-site detection. It is suitable for continuous water output status monitoring tasks under harsh conditions of submerged arc furnaces and is suitable for high-reliability positioning and recognition needs in scenarios with strong reflection, local overexposure, and variable fluid topology. Attached Figure Description

[0015] Figure 1 This is a flowchart of the visual inspection method for silicon water output from a submerged arc furnace in this invention.

[0016] Figure 2 This is the overall network architecture diagram of the visual inspection method for silicon water output from a submerged arc furnace in this invention.

[0017] Figure 3 This is a visualization of the feature space effect of t-SNE after fusing spatiotemporal features in this invention.

[0018] Figure 4This is a comparison diagram of the feature space before and after the introduction of the adversarial learning and discrimination module in this invention.

[0019] Figure 5 This is a display diagram of the video dataset of the effluent from the submerged arc furnace in this invention.

[0020] Figure 6 This is a statistical chart of the accuracy and intersection-over-union (IoU) ratio of the model in this invention after multiple tests.

[0021] Figure 7 This is a diagram illustrating the positioning and identification effect of the silicon water outlet status in the submerged arc furnace according to the present invention. Detailed Implementation

[0022] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Identical components are denoted by the same reference numerals. It should be noted that the terms "front," "rear," "left," "right," "upper," and "lower" used in the following description refer to directions in the accompanying drawings, and the terms "bottom surface," "top surface," "inner," and "outer" refer to directions toward or away from the geometric center of a specific component, respectively.

[0023] Due to the extreme high-brightness overexposure characteristics, severe obstruction by dense smoke and dust, drastic changes in industrial lighting, and varied fluid topology at the outlet of submerged arc furnaces, the acquired continuous video frames often suffer from severe overlap between static high-temperature backgrounds and dynamic fluids, blurred fluid edges, and susceptibility to false discontinuities caused by smoke interference. These problems make traditional visual inspection methods prone to misinterpreting static high-brightness furnace opening backgrounds as flowing molten silica, or misidentifying truly continuous molten silica fluid as a flow interruption due to smoke obstruction, resulting in inaccurate positioning and identification errors, severely limiting the accuracy and reliability of the inspection. Therefore, this invention provides a visual inspection method for molten silica outlet water from submerged arc furnaces, specifically as follows: Figure 1 As shown, it includes a dual-flow feature extraction step, a spatiotemporal feature modulation step, a fluid feature constraint step, a network joint optimization step, and a silicon water state detection step.

[0024] In the dual-stream feature extraction step, such as Figure 2As shown, continuous video frame images of a submerged arc furnace are acquired and deep features of appearance and motion are extracted separately using a dual-stream encoder. The dual-stream encoder is constructed based on a lightweight convolutional neural network architecture pre-trained on the ImageNet dataset. Its output feature map undergoes feature dimension compression and projection transformation via a lightweight channel adapter. The lightweight channel adapter is located at the output of the dual-stream encoder and enhances detailed features through a multi-scale module. This multi-scale module selectively aggregates feature maps from different layers of the encoder to construct a unified feature representation. Specifically, the input continuous video frame images are first uniformly scaled to a preset reference size, and their original time series are strictly maintained for paired or sequential input. This ensures that the motion branch can accurately extract fluid dynamic information between continuous frames while meeting the input requirements of the dual-stream encoder. Through these steps, dual-stream feature extraction achieves the extraction of multi-level spatiotemporal features from the original image.

[0025] In the spatiotemporal feature modulation step, the extracted deep motion features are used to calculate spatial weights through an attention mechanism, and the deep appearance features are dynamically modulated to form a fused feature. Specifically, the spatiotemporal feature modulation step includes a motion-gated modulation strategy, which involves generating spatial attention weights by activating the deep motion features, and multiplying them pixel-by-pixel with the deep appearance features to construct a fused spatiotemporal feature representation, the expression of which is: ; in, The output is the fused feature of the input image in the i-th frame. This is an element-wise multiplication operation between feature map channels. For activation function, For spatial feature convolution operation, For deep appearance features extracted from a dual-stream encoder, These are the deep motion features extracted from the dual-stream encoder.

[0026] In the fluid characteristic constraint step, a gravity continuity loss module is introduced into the feature space to calculate the gradient penalty in the vertical flow direction and generate the corresponding physical continuity constraint. Since real silica water exhibits a continuous downward flow under gravity, this step effectively overcomes the pseudo-flow interruption phenomenon caused by dense smoke obstruction by penalizing characteristic faults in the vertical direction. Specifically, the fluid characteristic constraint step also includes a gravity continuity loss function, which is used to calculate the characteristic distribution difference in the vertical flow direction to apply continuity constraints. Its expression is: ; in, To calculate the gravity coherence loss value, The preset gradient penalty function, Let (h, w) be the predicted output value of the feature map, and (h, w) be the spatial location. The predicted output value is generated for the feature of the position adjacent to (h, w) along the opposite direction of gravity. This generated physical coherence constraint is then passed to the next stage of network optimization.

[0027] In the network joint optimization step, such as Figure 3 As shown, the fused features are reconstructed using a lightweight codec, and the network is optimized by combining gravity coherence loss and multi-scale loss functions to enhance the physical consistency of fluid features and form good feature clustering in high-dimensional space. The lightweight codec construction strategy includes a backbone decoding path and a cross-layer skip connection path. The backbone decoding path consists of a feature reconstruction module composed of lightweight convolutional layers, batch normalization, and activation functions. The cross-layer skip connection path introduces a high-resolution feature concatenation layer. When optimizing the lightweight codec network, backpropagation is performed by minimizing the joint expected loss function on the training dataset, the expression of which is: ; Where L is the final optimized joint expectation loss function, and (X,Y) is the loss function derived from the training dataset. Mid-sampled input video frame image pairs and their corresponding real-world mask labels. This represents the total loss due to local coherence. To ensure the model's generalization ability, the submerged arc furnace effluent dataset used in this invention (see example) Figure 5 All images were collected from real production sites, covering various lighting conditions and heavy smoke obstruction. A total of over 4,822 valid image frames were extracted and labeled, and were strictly divided into training, validation and test sets in an 8:1:1 ratio.

[0028] Furthermore, the network joint optimization step includes a global reconstruction loss strategy, a local coherence loss strategy, and a dynamic feature discrimination and adversarial learning strategy. The global reconstruction loss strategy measures large-size fluid features by calculating the difference in the basic segmentation distribution, and its expression is: ; The local coherence loss strategy controls the effective area of ​​gravity penalty through dynamic region mask constraints, and its expression is: ;

[0029] Meanwhile, to distinguish between dynamic fluid features and static interference features caused by extreme brightness, a gating discrimination module was introduced during the network optimization stage. The specific effect of the feature space comparison before and after the introduction of this module is as follows: Figure 4As shown. This module updates network weights based on a hinged discriminant loss function with boundary margin, forcing the network to push apart the distribution boundary between the flowing silica water and the static background in the feature space. Its expression is:

[0030] ;

[0031] in, Here, is the dynamic feature discrimination loss function, where The probability distribution representing the dynamic characteristics of silicon water. The probability distribution representing the static highlighted background features. It is the output function of the gating discrimination module. This is the set boundary margin.

[0032] In the silica water state detection step, the optimized features are input into a segmentation network, and the outflow state of silica water is located and discriminated through connected component analysis and morphological filtering. The morphological and geometric constraint strategy includes independent connected component extraction of the initial feature mask, the expression of which is: ; Its arbitrary connected components Its features include an analytically computable geometric distribution: ; Where R is the set of extracted connected components, For connected component extraction operation, This serves as the initial binarization mask for the segmentation network output, where N is the total number of connected components. Let be the pixel area of ​​the k-th connected component. For spatial coordinates, The height of the bounding rectangle of the k-th connected region is given; then, the final silicon water positioning result is generated by combining the area and aspect ratio of the gravity-induced sagging shape with a joint threshold. ; in, To position the silica water as a mask after filtering out static noise. The preset minimum effective area threshold, Let the width of the bounding rectangle of the k-th connected component be . To prevent extremely small constants with a denominator of zero, This is the preset aspect ratio threshold.

[0033] Furthermore, the silica water state detection step also includes inference and effluent confidence assessment strategies. In the inference phase, such as... Figure 6The statistical charts of multiple tests of the model shown verify the high accuracy and robustness of the invention under complex working conditions. Specifically, the network output is upsampled and interpolated to the original video frame size and then Gaussian smoothed to obtain the pixel-level water discharge confidence score for accurate silica water positioning. The calculation process is as follows: ; in, Pixel-level water discharge confidence score. The initial probability score is the output of the segmentation network. and The Gaussian smoothing function and the upsampling interpolation function, respectively, are used to determine the frame-level water discharge state score for triggering system alarms. The maximum response value of all pixels in the array is given, and its expression is: ; in, The water discharge status score is at the frame level. Spatial coordinates The corresponding pixel-level output. The model calculation results are ultimately displayed through visualization tools as a localization and recognition effect diagram of the silicon water outlet status of the submerged arc furnace, as shown in the following figure. Figure 7 As shown.

[0034] To ensure that those skilled in the art can implement and reproduce the model training process of this invention, the model training configuration and hyperparameter settings for the network joint optimization step are as follows: The network model training process is performed within a deep learning framework, using the AdamW optimizer for joint updates of network parameters, wherein the weight decay coefficient is set to... Momentum parameters and Set them to 0.9 and 0.999 respectively; set the initial learning rate to... Furthermore, a cosine annealing strategy was employed during training to dynamically decay the learning rate, ensuring the model's convergence stability in the later stages of training. Additionally, considering the memory limitations of edge computing devices and the dimensionality requirements of dual-stream feature inputs, the batch size for model training was set to 8, and the total number of training epochs was set to 200. For the multiple loss functions in the network joint optimization, the global reconstruction loss weights were experimentally calibrated. and The local gravity constraint strength coefficients are set to 1.0 and 0.5. Set to 0.2.

[0035] Meanwhile, to further verify the effectiveness of the technical solution of this invention, a comparative experiment was conducted between the method of this invention and the traditional background subtraction method and the basic single-stream static segmentation model (standard U-Net) under the same test conditions. The results are as follows: (1) For static high light interference: In the test under strong light conditions, the false detection rate of the traditional differential method and the basic single flow model is about 35% and 12%, respectively. This invention introduces a spatiotemporal feature modulation and dynamic feature discrimination module, which increases the distinction between dynamic fluid and static background in the feature space, thereby reducing the false detection rate of the high-brightness area to about 1.5%.

[0036] (2) Regarding the phenomenon of dense smoke obstruction and flow interruption: In the environment of dense smoke obstruction, the basic single-flow model is prone to predictive morphological breaks, and the false flow interruption miss rate is about 18%. The fluid characteristic constraint step of the present invention applies continuity constraints through gravity continuity loss, which reduces the morphological breaks caused by obstruction, controls the flow interruption miss rate to within 3%, and improves the overall intersection-to-union ratio (IoU) by about 8.5% compared with the above basic model.

[0037] (3) Regarding the matching degree between computational efficiency and application: The present invention adopts a lightweight codec structure, and the single-frame inference time is kept within 50 milliseconds on edge computing devices with limited computing power. This processing efficiency can meet the production cycle requirements of real-time detection in electric arc furnaces, and achieves a good balance between detection accuracy and computational resource consumption.

[0038] The visual inspection method for molten silicon outlet water from submerged arc furnaces proposed in this invention has been successfully applied to the intelligent transformation project of submerged arc furnaces in a steel group. Since the system was deployed and put into operation, it has been running stably for more than three months under real-world conditions such as extreme high temperatures at the furnace mouth, overexposure to strong light, and obstruction by dense smoke and dust. Actual industrial application data shows that this method effectively overcomes the problem of traditional visual inspection methods being susceptible to interference from static high light and false tomography. Under various complex operating conditions, the overall accuracy of molten silicon state positioning and identification remains above 97%, and even under comprehensive extreme operating conditions, the intersection-to-union ratio (IoU) remains stable above 90%. At the same time, while ensuring high accuracy, the algorithm processing efficiency of this inspection system meets the real-time monitoring requirements of industrial sites, realizing all-weather automated monitoring, eliminating the reliance on high-intensity manual monitoring in traditional smelting processes, and significantly improving the production safety and intelligence level of the submerged arc furnace outlet water stage.

[0039] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for visually inspecting molten silica water from a submerged arc furnace, characterized in that: Includes the following steps: The dual-stream feature extraction step acquires continuous video frame images of the submerged arc furnace, extracts deep appearance features and deep motion features respectively through a dual-stream encoder, and enhances detailed features through a multi-scale module; The spatiotemporal feature modulation step involves calculating spatial weights for the extracted motion deep features through an attention mechanism, and then dynamically modulating the appearance deep features to form fused features. The fluid feature constraint step introduces a gravity coherence loss module into the feature space to calculate the gradient penalty in the vertical flow direction and generate the corresponding physical coherence constraint. The network joint optimization step reconstructs the fused features using a lightweight codec, and optimizes the network by combining global reconstruction loss, gravity coherence loss, multi-scale loss function, and dynamic feature discrimination loss to enhance the physical consistency of fluid features. The silicon water state detection step involves inputting the optimized features into a segmentation network, and using connected component analysis and morphological filtering to locate and determine the outflow state of the silicon water.

2. The visual inspection method for silicon molten metal outlet water from a submerged arc furnace according to claim 1, characterized in that: The dual-stream encoder in the dual-stream feature extraction step is built on a lightweight convolutional neural network architecture pre-trained on the ImageNet dataset. Its output feature map is compressed and projected through a lightweight channel adapter.

3. The visual inspection method for silicon molten metal outlet water from a submerged arc furnace according to claim 1, characterized in that: The spatiotemporal feature modulation step includes a motion-gated modulation strategy. This strategy involves generating spatial attention weights by activating deep motion features, and then multiplying these weights pixel-by-pixel with the deep appearance features to construct a fused spatiotemporal feature representation, expressed as: ； in, The output is the fused feature of the input image in the i-th frame. This is an element-wise multiplication operation between feature map channels. For activation function, For spatial feature convolution operation, For deep appearance features extracted from a dual-stream encoder, These are the deep motion features extracted from the dual-stream encoder.

4. The visual inspection method for silicon molten metal outlet water from a submerged arc furnace according to claim 3, characterized in that: The fluid characteristic constraint step also includes a gravity continuity loss function, which is used to calculate the characteristic distribution difference in the vertical flow direction to apply continuity constraints. Its expression is as follows: ； in, To calculate the gravity coherence loss value, The preset gradient penalty function, Let (h, w) be the predicted output value of the feature map, and (h, w) be the spatial location. The predicted output value is the feature value of the position adjacent to (h,w) along the opposite direction of gravity.

5. A visual inspection method for silicon molten metal discharge from a submerged arc furnace according to claim 1 or 4, characterized in that: The silica water state detection step includes morphological and geometric constraint strategies. These strategies include extracting independent connected components from the initial feature mask, expressed as follows: ； Its arbitrary connected components Its features include an analytically computable geometric distribution: ； Where R is the set of extracted connected components, For connected component extraction operation, This serves as the initial binarization mask for the segmentation network output, where N is the total number of connected components. Let be the pixel area of ​​the k-th connected component. For spatial coordinates, The height of the bounding rectangle of the k-th connected region is given; then, the final silicon water positioning result is generated by combining the area and aspect ratio of the gravity-induced sagging shape with a joint threshold. ； in, To position the silica water as a mask after filtering out static noise. The preset minimum effective area threshold, Let the width of the bounding rectangle of the k-th connected component be . To prevent extremely small constants with a denominator of zero, This is the preset aspect ratio threshold.

6. A visual inspection method for silicon molten metal discharge from a submerged arc furnace according to claim 1 or 4, characterized in that: The network joint optimization step includes a global reconstruction loss strategy and a local coherence loss strategy. The global reconstruction loss strategy includes measuring large-size fluid characteristics by calculating the difference in the basic segmentation distribution, and its expression is: ； The local coherence loss strategy controls the effective area of ​​gravity penalty through dynamic region mask constraints, and its expression is as follows: ； in, For the overall reconstruction loss, Let P be the total loss for local coherence, P be the network's predicted output features, and Y be the true state label. For adaptive constraint binary mask, The gravitational constraint strength coefficient. and These are the dynamic weighting coefficients corresponding to the global loss. This is due to the loss of local gravity continuity.

7. The visual inspection method for silicon molten metal outlet water from a submerged arc furnace according to claim 6, characterized in that: The network joint optimization step includes a lightweight encoder-decoder construction strategy, which comprises a backbone decoding path and a cross-layer skip connection path. The backbone decoding path consists of a feature reconstruction module composed of lightweight convolutional layers, batch normalization, and activation functions. The cross-layer skip connection path introduces a high-resolution feature concatenation layer. When optimizing the lightweight encoder-decoder network, backpropagation is performed by minimizing the joint expected loss function on the training dataset, the expression of which is: ； Where L is the final optimized joint expectation loss function, and (X,Y) is the loss function derived from the training dataset. Mid-sampled input video frame image pairs and their corresponding real-world mask labels. This represents the total loss due to local continuity issues.

8. The visual inspection method for silicon molten metal outlet water from a submerged arc furnace according to claim 1, characterized in that: The network joint optimization step also includes a dynamic feature discrimination and adversarial learning strategy. The adversarial learning strategy includes constructing a gated discrimination module. This gated discrimination module updates network weights based on a hinged discrimination loss function with boundary margin to distinguish between dynamic fluid features and static interference features. Its expression is: ； in, Here, is the dynamic feature discrimination loss function, where The probability distribution representing the dynamic characteristics of silicon water. The probability distribution representing the static highlighted background features. It is the output function of the gating discrimination module. The set boundary margin; through the adversarial learning strategy, the network is forced to push away the distribution boundary between the flowing silicon water and the static background in the feature space.

9. The visual inspection method for silicon molten metal outlet water from a submerged arc furnace according to claim 8, characterized in that: The silica water state detection step also includes an inference and water outflow confidence assessment strategy. This strategy involves upsampling and interpolating the network output to the original video frame size and then smoothing it using Gaussian smoothing to obtain a pixel-level water outflow confidence score for precise silica water localization. The calculation process is as follows: ； in, The pixel-level water discharge confidence score is used to trigger system alarms, while the frame-level water discharge status score is determined by... The maximum response value of all pixels in the array is given, and its expression is: ； in, The water discharge status score is at the frame level. The initial probability score is the output of the segmentation network. and These are the Gaussian smoothing function and the upsampling interpolation function, respectively. Spatial coordinates The corresponding pixel-level output.

Images