Three-field synergy-based bubble suppression system and method for borosilicate glass all-electric melting furnace

Abstract

The application discloses a borosilicate glass full-electric melting furnace bubble suppression system and method based on three-field cooperation and relates to the technical field of industrial control. Discharge data of a glass melting furnace is acquired to obtain first characteristic data; temperature data of the glass melting furnace is acquired to construct a temperature field and obtain second characteristic data; a fluid image of the surface of the glass melting furnace is acquired, texture features of the fluid surface are identified through a target model, the texture features are added to the fluid image to obtain an enhanced image, a two-dimensional velocity vector diagram is calculated, and third characteristic data is obtained by performing feature extraction on the two-dimensional velocity vector diagram; a parameter abnormality type is determined, and bubble suppression is performed. By fusing multi-source data of discharge, temperature and fluid images, three-field features are extracted, and the abnormality type is determined, so that bubble causes can be accurately identified, fine management and control of working conditions can be realized, bubble generation, retention and regeneration problems can be reduced, and the accuracy and automation level of bubble suppression in the glass melting process can be improved.

Description

Technical Field

[0001] This invention relates to the field of industrial control technology, specifically to a bubble suppression system and method for a borosilicate glass all-electric melting furnace based on three-field synergy. Background Technology

[0002] Borosilicate glass, with its excellent heat resistance and chemical stability, continues to see its applications expand. All-electric melting furnaces, due to their high thermal efficiency and flexible temperature control, have become the mainstream melting equipment for borosilicate glass. However, during the operation of these furnaces, air bubbles are prone to form inside the molten glass, directly affecting the quality of the finished product. Currently, the mainstream bubble control methods in the industry have many limitations. Existing monitoring methods mostly use single-point thermocouples and conventional meters for data collection, which can only obtain electric field and temperature information at local points in the furnace, failing to form a comprehensive monitoring system. For the flow field of molten glass inside the furnace, most methods rely on manual observation of the liquid surface, making it difficult to capture details such as minute surface textures, hidden eddies, and flow stagnation.

[0003] Existing methods for suppressing bubbles in all-electric borosilicate glass melting furnaces mostly rely on manual experience to judge operating conditions. They cannot conduct comprehensive and quantitative monitoring and analysis of the three core parameters: electric field, temperature field, and flow field. This makes it difficult to accurately locate the abnormal operating conditions corresponding to bubble generation, resulting in a lack of targeted bubble suppression measures. Summary of the Invention

[0004] The purpose of this invention is to address the problem mentioned in the background art that relies heavily on human experience to judge operating conditions, making it impossible to conduct comprehensive and quantitative monitoring and analysis of the three core parameters of electric field, temperature field, and flow field, making it difficult to accurately locate the abnormal operating conditions corresponding to bubble generation, resulting in a lack of targeted bubble suppression measures. Therefore, this invention proposes a bubble suppression system and method for borosilicate glass all-electric melting furnaces based on the synergy of the three fields.

[0005] A first aspect of this invention provides a method for suppressing bubbles in a fully electric melting furnace for borosilicate glass based on three-field synergy, the method comprising:

[0006] The discharge data of the glass furnace is acquired, and the first feature data is obtained by extracting features from the discharge data;

[0007] Temperature data of a glass melting furnace is acquired, a temperature field is constructed based on the temperature data, and features are extracted from the temperature field to obtain second feature data; the temperature data includes: outer side data and inner side data;

[0008] High-definition infrared cameras are used to capture fluid images on the surface of glass furnaces. The texture features of the fluid surface are identified using a target model, and these texture features are then added to the fluid images to obtain enhanced images.

[0009] The enhanced image is calculated using optical flow to obtain a two-dimensional velocity vector map, and the third feature data is obtained by feature extraction from the two-dimensional velocity vector map.

[0010] The parameter anomaly type is determined based on the first feature data, the second feature data, and the third feature data, and bubble suppression is performed based on the parameter anomaly type.

[0011] By fusing multi-source data from discharge, temperature, and fluid images, extracting features from the three fields, and determining the type of anomaly, the causes of bubbles can be accurately identified, enabling refined control of operating conditions, reducing problems related to bubble generation, retention, and regeneration, and improving the accuracy and automation level of bubble suppression during glass melting.

[0012] Optionally, a temperature field is constructed based on the temperature data, including:

[0013] The outer data and the inner data are preprocessed to obtain preprocessed outer data and preprocessed inner data, respectively. A three-dimensional model of the glass furnace is obtained, and the three-dimensional model is divided to obtain a spatial mesh model.

[0014] The preprocessed inner data is added to the spatial grid model to obtain the point assignment grid, and the preprocessed outer data is added to the point assignment grid according to the temperature decay law to obtain the global grid;

[0015] The empty grid in the global grid is filled with data by spatial interpolation algorithm to obtain complete grid data, and the temperature field is obtained by spatial mapping of the complete grid data.

[0016] By integrating temperature data from both inside and outside the furnace, and using a 3D model of the furnace and a spatial grid to assign data in layers, and combining temperature decay patterns with spatial interpolation to fill in missing data, the system can fill in the monitoring blind spots of single-point temperature measurement, fully restore the temperature distribution throughout the furnace, and construct an accurate three-dimensional temperature field. This provides comprehensive and reliable data support for subsequent temperature field feature extraction and anomaly detection.

[0017] Optionally, before identifying the texture features of the fluid surface through the target model, the training process of the target model includes:

[0018] A set of raw fluid infrared images of the surface of a glass furnace is acquired by a high-definition infrared camera. The raw fluid infrared images are then manually labeled to obtain a set of artificially labeled images. The artificially labeled images are then added to the raw fluid infrared images to obtain a training image set.

[0019] The training image set is input into the target model to obtain the confidence level of the texture features. If the confidence level is greater than the threshold, the target model is deemed to have passed the training.

[0020] The collected infrared images, combined with manual annotations, form a training image set that closely matches the actual morphological features of the glass melt surface texture, ensuring the authenticity and suitability of the training samples. Using feature recognition confidence as the evaluation criterion to complete model training can effectively improve the accuracy and reliability of the target model in recognizing liquid surface texture features, providing an accurate image foundation for subsequent image enhancement, flow field calculation, and feature extraction.

[0021] Optionally, the target model is an improvement based on the YOLO11 model, including:

[0022] The target model is obtained by replacing the backbone network with an improved backbone network and the neck structure with an improved neck structure; the YOLO11 model includes a backbone network and a neck structure.

[0023] Optionally, the improved backbone network operation process includes:

[0024] Obtain the original feature map A, input the original feature map A into the SRFD module to obtain the first feature map A, input the first feature map A into the CSFPA module, DRFD module and CSFPA module in sequence to obtain the second feature map A, input the second feature map A into the DRFD module and CSFPA module in sequence to obtain the third feature map A, input the third feature map A into the DRFD module, CSFPA module, SPPF module and C2PSA module in sequence to obtain the fourth feature map A, and use the second feature map A, the third feature map A and the fourth feature map A as the output of the improved backbone network.

[0025] Based on the improved backbone network described above, and combining the SRFD module for fidelity downsampling of shallow texture details, the DRFD module for stable transmission of deep semantics, the CSPFA module for multi-scale grouped convolutional aggregation, and the MSW module for morphology-spectrum-window joint gating modulation in lateral connections, this model can suppress background interference caused by periodic textures and specular highlights in the task of extracting texture features from molten glass surfaces, while enhancing the edge response and structural coherence of small, low-contrast defects. Through the synergistic complementarity of each module, the model improves the separability of traps from the background while maintaining a lightweight structure, reducing false detections and false negatives caused by texture confusion, thereby achieving high-precision and robust feature representation.

[0026] Optionally, the process of improving the neck structure includes:

[0027] The 4th feature map A is upsampled to obtain the 5th feature map A. The 5th feature map A is concatenated with the 3rd feature map A to obtain the 6th feature map A. The 6th feature map A is input into the CSFPA module to obtain the 7th feature map A. The 7th feature map A is upsampled and concatenated with the 2nd feature map A to obtain the 8th feature map A. The 8th feature map A is input into the CSFPA module to obtain the 9th feature map A.

[0028] The 9th feature map A is input into the DRFD module and then concatenated with the 7th feature map A to obtain the 10th feature map A. The 10th feature map A is input into the CSFPA module to obtain the 11th feature map A. The 11th feature map A is input into the DRFD module and then concatenated with the 4th feature map A to obtain the 12th feature map A. The 12th feature map A is input into the CSFPA module to obtain the 13th feature map A. The 7th feature map A, the 11th feature map A, and the 13th feature map A are used as the output of the improved neck structure.

[0029] By progressively upsampling deep feature maps and then stitching them together with corresponding shallow feature maps across scales, followed by multi-scale grouped convolutional aggregation via the CSPFA module and stable downsampling via the DRFD module before fusion, a dense feature transfer path with bidirectional interaction from top to bottom and bottom to top is formed. This design can fully reuse feature information at different resolutions, effectively enhancing the ability to preserve details of small, low-contrast defects in glass melt surface texture feature extraction tasks. Simultaneously, through multiple rounds of aggregation and modulation of multi-level fused features, it suppresses background responses caused by texture repetition and uneven illumination, making defect boundaries clearer and feature representations more continuous. This improves the adaptability and discriminative power of the multi-scale features output from the neck section to complex surfaces, providing a more stable and discriminative feature representation for the detection head.

[0030] Optionally, a two-dimensional velocity vector map is obtained by calculating the enhanced image using optical flow, including:

[0031] Feature point set is obtained by extracting features from the enhanced image, and inter-frame displacement dataset is obtained by performing optical flow calculation on two temporally adjacent feature coordinate point sets; the feature points include: liquid surface texture inflection points and wave contour points;

[0032] The feature point velocity vector set is calculated based on the inter-frame displacement dataset. The enhanced image is divided into regions to obtain a uniform computational grid. The feature point velocity vectors in the feature point velocity vector set are added to the uniform computational grid to obtain grid vector data. The grid vector data is padded to obtain complete grid vector data. The complete grid vector data is rendered to obtain a two-dimensional velocity vector map.

[0033] By extracting inflection points of liquid surface texture and undulation contour points from enhanced images as feature point sets, and calculating the velocity vectors of these feature points using displacement information between adjacent frames, and then mapping them onto a uniformly divided computational grid and performing incomplete rendering, discrete local motion information can be transformed into a globally continuous, spatially complete two-dimensional velocity vector map. In the task of extracting surface texture features from molten glass, this method effectively captures the dynamic changes of minute fluctuations and texture deformations on the liquid surface, clearly presenting the direction, intensity, and spatial distribution characteristics of molten flow, providing intuitive and quantitative motion information support for subsequent analysis of surface stability, defect evolution, or flow field anomalies.

[0034] A second aspect of this invention provides a bubble suppression system for a borosilicate glass all-electric melting furnace based on three-field synergy, the system comprising:

[0035] The discharge data acquisition module is used to acquire discharge data from the glass furnace and extract features from the discharge data to obtain the first feature data.

[0036] A temperature data acquisition module is used to acquire temperature data of a glass melting furnace, construct a temperature field based on the temperature data, and extract features from the temperature field to obtain second feature data; the temperature data includes: outer side data and inner side data;

[0037] The flow field feature acquisition module is used to capture fluid images on the surface of the glass furnace using a high-definition infrared camera, identify the texture features of the fluid surface through a target model, and add the texture features to the fluid image to obtain an enhanced image;

[0038] The flow field feature extraction module is used to calculate the enhanced image using optical flow to obtain a two-dimensional velocity vector map, and to extract features from the two-dimensional velocity vector map to obtain the third feature data.

[0039] The bubble suppression module is used to determine the parameter anomaly type based on the first feature data, the second feature data, and the third feature data, and to suppress bubbles based on the parameter anomaly type. Attached Figure Description

[0040] Figure 1 A flowchart of a method for suppressing bubbles in a fully electric melting furnace for borosilicate glass based on three-field synergy, provided in an embodiment of the present invention;

[0041] Figure 2 A schematic diagram of the model structure of the borosilicate glass all-electric melting furnace bubble suppression method based on three-field synergy provided in the embodiments of the present invention. Detailed Implementation

[0042] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0043] The present invention provides a method for suppressing bubbles in a fully electric melting furnace for borosilicate glass based on three-field synergy. See also Figure 1 , Figure 1 A flowchart illustrating a method for suppressing bubbles in a fully electric melting furnace for borosilicate glass based on three-field synergy, provided in an embodiment of the present invention. The method includes the following steps:

[0044] S101, Obtain the discharge data of the glass furnace, and extract the first feature data from the discharge data;

[0045] S102, acquire the temperature data of the glass furnace, construct a temperature field based on the temperature data, and extract features from the temperature field to obtain the second feature data;

[0046] S103 captures fluid images of the glass furnace surface using a high-definition infrared camera, identifies the texture features of the fluid surface using a target model, and adds the texture features to the fluid image to obtain an enhanced image;

[0047] S104, the enhanced image is calculated by optical flow method to obtain a two-dimensional velocity vector map, and the feature data of the two-dimensional velocity vector map is obtained by feature extraction.

[0048] S105, determine the parameter anomaly type based on the first feature data, the second feature data and the third feature data, and perform bubble suppression based on the parameter anomaly type.

[0049] Temperature data includes: outer and inner data;

[0050] In one implementation, a borosilicate glass product (sample) produced has been found to have product defects, with too many bubbles. Therefore, it is necessary to reverse-engineer the production line and determine the cause of the bubbles.

[0051] In one implementation, glass furnace discharge data (voltage, current, power, discharge pulses, and conductivity / resistivity, etc.) are collected collaboratively by in-furnace electrodes, a high-frequency current transformer, a voltage transmitter, a high-speed data acquisition card, and a DCS / edge computing unit. In the all-electric furnace, platinum / molybdenum electrodes are inserted into the molten glass to form a discharge circuit. A high-frequency current transformer (20kHz–300kHz) is used to pick up the current pulses of the electrode branches, and a high-voltage voltage transmitter synchronously acquires the voltage waveforms between the electrodes. A high-speed acquisition card (sampling rate ≥1MS / s) is used to capture transient discharge waveforms, and an LCR meter is used to measure the equivalent resistance / conductivity of the molten liquid. After signal filtering, amplification, and analog-to-digital conversion, the data is summarized, preprocessed, and stored in real time by the DCS system or edge computing unit, outputting a discharge dataset containing discharge frequency, peak voltage, pulse current, discharge energy, and impedance characteristics for subsequent feature extraction.

[0052] In one implementation, the collected furnace discharge dataset is preprocessed by noise reduction and normalization. Then, feature extraction is performed in the time and frequency domains. In the time domain, the average voltage, peak current, power fluctuation amplitude, current density, waveform distortion rate, and discharge duration are extracted. In the frequency domain, the spectral energy, harmonic components, and frequency fluctuation features are extracted. At the same time, the current deviation value and power imbalance of each electrode circuit are statistically analyzed. The above quantitative indicators are integrated into the first feature data characterizing the electric field operation status, thereby reflecting the uniformity and stability of the electric field in the furnace and providing a basis for subsequent judgment of bubble induction and implementation of suppression strategies.

[0053] In one implementation, after preprocessing the three-dimensional temperature field of the furnace reconstructed from the fused internal and external temperature measurement data by grid data normalization and outlier removal, the core features of the temperature field are extracted: the average temperature, extreme temperature, regional temperature difference, longitudinal and transverse temperature gradients, temperature fluctuation variance, area of ​​high-temperature hot spots, and isothermal stability of the melting zone, clarifying zone, and forming zone inside the furnace are calculated. The parameters of each temperature dimension are uniformly encoded and integrated by a weighted summation formula to obtain a set of quantitative parameters that includes the temperature amplitude of the entire furnace, temperature distribution uniformity, temperature gradient change law, and temperature time-series fluctuation characteristics. This set of quantitative parameters is defined as the second feature data, which is used to characterize the stable state and abnormal defects of the three-dimensional temperature field inside the glass furnace.

[0054] In one implementation, after preprocessing the two-dimensional velocity vector diagram of the glass melt surface obtained by optical flow method through mesh normalization and invalid vector removal, the core features of the flow field are extracted in the whole domain: the mean velocity, extreme velocity, velocity fluctuation variance, flow direction angle, local eddy intensity, flow field uniformity, area of ​​liquid surface flow stagnation region and inter-frame flow stability coefficient are calculated for each region of the melt surface. The parameters of each flow field dimension are uniformly encoded and fused by weighted summation formula to obtain a set of quantitative parameters including the magnitude of the flow velocity on the melt surface, the distribution of flow direction, the degree of flow field turbulence, the characteristics of eddy defects and the temporal flow change law. This set of quantitative parameters is defined as the third feature data, which is used to accurately characterize the steady state and flow abnormal defects of the glass furnace melt flow field.

[0055] In one implementation, by combining monitoring data from three fields (electric field, temperature field, and flow field), the cause of bubble formation is located. After detecting the data from the three fields, the bubble fault type is matched against the normal process range.

[0056] Abnormal electric field data leads to numerous bubbles (localized overheating type bubbles):

[0057] Normal range: The current density in each zone is uniform, the power fluctuation is within ±2%, and there is no bias current.

[0058] Abnormal manifestations: single electrode current is too high / too low, local power rises and falls suddenly, and current distribution around the electrode is uneven.

[0059] Corresponding bubble characteristics: A large number of fine secondary bubbles appear near the electrodes and at the bottom of the kiln.

[0060] An imbalance in the electric field leads to localized high-temperature gas evolution, which must be suppressed by adjusting the electric field.

[0061] Abnormal temperature field data leads to numerous bubbles (difficulty in rising / secondary bubbles):

[0062] Normal range: The overall temperature gradient is stable with the bottom heating and the top cooling; the temperature fluctuation in the clarification zone is ≤±3℃; there is no sudden cooling or heating throughout the entire process.

[0063] Abnormal phenomena: Local temperatures far exceed the set value (hot spots); low temperature in the clarification zone and high viscosity of the molten glass; excessively rapid cooling in the discharge section and sudden temperature changes.

[0064] Corresponding bubble characteristics: bubbles remain suspended for a long time and have difficulty rising; a large number of small bubbles suddenly appear at the discharge end;

[0065] The temperature field is disordered, and the temperature field must be adjusted to suppress bubbles.

[0066] Abnormal flow field data leads to an abundance of bubbles (trapped / entrained bubbles):

[0067] Normal range: The liquid flow is a stable laminar flow; the flow velocity is uniform; there are no obvious eddies or backflows; the residence time of the glass melt in the clarification zone meets the standard.

[0068] Abnormal signs: Localized flow velocity too fast / too slow; eddies and backflow detected behind flow channels and dams; large fluctuations in liquid level.

[0069] Corresponding bubble characteristics: Bubbles are trapped in dead corners and repeatedly circulate; there are many bubbles near the liquid surface (air is entrained).

[0070] Turbulent flow field necessitates optimization to suppress bubbles.

[0071] In one implementation, an abnormal electric field induces bubbles (local overheating, secondary bubbles). The adjustment objectives are: to achieve uniform current / power, eliminate hot spots, and prevent additional gas generation.

[0072] Electrode power zone fine-tuning: Reduce the output power of electrode groups with excessive current or temperature; slightly increase the power in areas with insufficient heating to even out the heat across the entire area;

[0073] Balanced current distribution: Check electrode wiring and power supply circuit, and eliminate line faults; if the electrodes are aged or have accumulated material, stop the machine to clean / replace them to avoid uneven current distribution.

[0074] Tighten the power fluctuation threshold: Optimize the parameters of the electronic control system to limit power fluctuations to a smaller range and prevent frequent temperature fluctuations from generating bubbles;

[0075] Abnormal temperature field causes bubbles (high viscosity, slow rise, bubbles generated by sudden cooling). Adjustment goals: repair the temperature gradient, clarify at a constant temperature, and cool down gradually.

[0076] Local hotspots exist: By combining electric field synchronous power reduction, the abnormally high temperature zone is reduced to the process range, and gas evolution is suppressed;

[0077] Low temperature in the clarification zone makes it difficult for bubbles to rise: Slightly increase the overall temperature of the clarification zone, reduce the viscosity of the molten glass, and reduce the resistance to bubble rise; strictly maintain a constant temperature and prohibit frequent temperature changes;

[0078] Rapid cooling in the discharge section (secondary bubbles): Slow down the cooling rate, widen the cooling range, allow the gas to dissolve slowly, and avoid sudden temperature drop and bubble formation.

[0079] Optimize the vertical temperature gradient: ensure the highest temperature at the bottom of the kiln and gradually decrease it upwards, enhance natural convection, and drive the bubbles to rise.

[0080] Abnormal flow field causes bubbles (vortex retention, air entrainment, insufficient residence time). Adjustment goal: form stable laminar flow, eliminate vortices, and prolong clarification time;

[0081] Large fluctuations in liquid level and severe air entrapment: Stabilize feeding and discharging speeds to maintain a constant liquid level; reduce the impact of feeding and discharging.

[0082] If eddies / backflow (bubbles trapped) are detected, fine-tune the bubbling system: adjust the bubbling volume and air inlet position, change the local flow direction, and disperse the eddies;

[0083] Inspect the dam and flow tunnel: if the structure is worn or has accumulated slag, clean and repair it in time, and force the liquid to flow along the preset path;

[0084] Flow rate too fast, clarification time insufficient: reduce the overall discharge flow rate, extend the residence time of the glass melt in the clarification zone, and give the bubbles enough time to rise and be discharged.

[0085] If the flow rate is too slow in some areas and dead zones are formed: slightly increase the temperature field, enhance natural flow, or increase the amount of bubbling in the corresponding area to activate the stagnant liquid flow.

[0086] In one embodiment, constructing a temperature field based on temperature data includes:

[0087] The outer and inner data are preprocessed to obtain preprocessed outer data and preprocessed inner data, respectively. A three-dimensional model of the glass furnace is obtained, and the three-dimensional model is divided to obtain a spatial mesh model.

[0088] The preprocessed inner data is added to the spatial grid model to obtain the point assignment grid, and the preprocessed outer data is added to the point assignment grid according to the temperature decay law to obtain the global grid;

[0089] The empty grid in the global grid is filled with data by spatial interpolation algorithm to obtain complete grid data, and the temperature field is obtained by spatial mapping of the complete grid data.

[0090] In one implementation, an infrared camera is used to collect temperature data of the furnace wall and liquid surface area of ​​the glass melting furnace to obtain the outer side data; thermocouples installed inside the furnace are used to collect temperature data of different depths and different areas of the melting furnace to obtain the inner side data.

[0091] In one implementation, data cleaning is performed on the outer and inner data respectively to remove abnormal jump values ​​and missing values, resulting in preprocessed outer and inner data; combined with the three-dimensional geometric model of the furnace, the furnace space is divided into three-dimensional meshes, and uniform spatial mesh units are divided to generate a spatial mesh model.

[0092] In one implementation, the preprocessed inner data is mapped one by one to the corresponding grid cells according to the thermocouple installation coordinates to complete the temperature assignment of discrete points in the furnace and obtain the point assignment grid.

[0093] By preprocessing the external data and combining the heat conduction model and the thermal resistance parameters of the furnace body refractory material, the temperature decay law of the furnace wall from the outside to the inside is calculated, and the temperature distribution function of the furnace wall region is obtained. Using the temperature distribution function, the blank grid cells around the furnace wall are filled with temperature to obtain the global grid.

[0094] In one implementation, based on a spatial interpolation algorithm, the blank grid inside the furnace without sensors is interpolated using the assigned grid points as a reference to complete the temperature values ​​of all grid units, thereby obtaining complete grid data. The complete grid data is then spatially mapped and visualized to generate a three-dimensional temperature field.

[0095] In one embodiment, the texture features of a fluid surface are identified through a target model. The process of identifying the target model includes:

[0096] Obtain the original feature map A, input the original feature map A into the SRFD module to obtain the first feature map A, input the first feature map A into the CSFPA module, DRFD module and CSFPA module in sequence to obtain the second feature map A, input the second feature map A into the DRFD module and CSFPA module in sequence to obtain the third feature map A, input the third feature map A into the DRFD module, CSFPA module, SPPF module and C2PSA module in sequence to obtain the fourth feature map A, and use the second feature map A, the third feature map A and the fourth feature map A as the output of the improved backbone network;

[0097] The 4th feature map A is upsampled to obtain the 5th feature map A. The 5th feature map A is concatenated with the 3rd feature map A to obtain the 6th feature map A. The 6th feature map A is input into the CSFPA module to obtain the 7th feature map A. The 7th feature map A is upsampled and concatenated with the 2nd feature map A to obtain the 8th feature map A. The 8th feature map A is input into the CSFPA module to obtain the 9th feature map A.

[0098] The 9th feature map A is input into the DRFD module and then concatenated with the 7th feature map A to obtain the 10th feature map A. The 10th feature map A is input into the CSFPA module to obtain the 11th feature map A. The 11th feature map A is input into the DRFD module and then concatenated with the 4th feature map A to obtain the 12th feature map A. The 12th feature map A is input into the CSFPA module to obtain the 13th feature map A. The 7th feature map A, the 11th feature map A, and the 13th feature map A are used as the output for improving the neck structure.

[0099] In one implementation, the working principle of the SRFD module includes:

[0100] Obtain the original tensor A, input the original tensor A into the Conv module to obtain the first tensor A, input the first tensor A into the GConv module and the DWConv module in sequence to obtain the second tensor A, and input the first tensor A into the Conv module after downsampling to obtain the third tensor A.

[0101] The second tensor A is concatenated with the third tensor A to obtain the fourth tensor A. The fourth tensor A is downsampled to obtain the fifth tensor A. The fourth tensor A is input into the GConv module to obtain the sixth tensor A. The sixth tensor A is max-pooled to obtain the seventh tensor A. The sixth tensor A is input into the DWConv module to obtain the eighth tensor A. The fifth, seventh, and eighth tensors A are concatenated to obtain the ninth tensor A. The ninth tensor A is used as the output of the SRFD module.

[0102] In one implementation, the working principle of the DRFD module includes:

[0103] Obtain the original tensor B, downsample the original tensor B to obtain the first tensor B, input the original tensor B into the GConv module to obtain the second tensor B, input the second tensor into the DWConv module to obtain the third tensor B, perform max pooling on the second tensor B to obtain the fourth tensor B, concatenate the first tensor B, the third tensor B and the fourth tensor B to obtain the fifth tensor B, and use the fifth tensor B as the output of the DRFD module.

[0104] In one implementation, the working principle of the CSPFA module includes:

[0105] Obtain the original tensor C, input the original tensor C into the Conv module to obtain the first tensor C, split the first tensor C to obtain the first branch tensor C and the second branch tensor C, input the second branch tensor C into the PFA module to obtain the third branch tensor C, input the third branch tensor C into the PFA module to obtain the fourth branch tensor C, concatenate the first branch tensor C, the second branch tensor C, the third branch tensor C and the fourth branch tensor C to obtain the second tensor C, input the second tensor C into the Conv module to obtain the third tensor C, and use the third tensor C as the output of the CSPFA module.

[0106] In one embodiment, the target model is based on an improvement of the YOLO11 model, including:

[0107] The target model is obtained by replacing the backbone network with an improved backbone network and the neck structure with an improved neck structure; the YOLO11 model includes a backbone network and a neck structure.

[0108] In one implementation, see [link to implementation details]. Figure 2 , Figure 2A schematic diagram of the model structure of the borosilicate glass all-electric melting furnace bubble suppression method based on three-field synergy provided in the embodiments of the present invention.

[0109] In one embodiment, the training process of a target model for identifying the texture features of a fluid surface includes:

[0110] The original fluid infrared image set on the surface of the glass furnace is acquired by a high-definition infrared camera. The original fluid infrared image set is identified and labeled by humans to obtain a manually labeled set. The manually labeled set is added to the original fluid infrared image set to obtain a training image set.

[0111] The confidence score of the texture features is obtained by inputting the training image set into the target model. If the confidence score is greater than the threshold, the target model is considered to have been trained successfully.

[0112] In one implementation, a high-definition infrared camera acquires an original fluid infrared image of the surface of a glass furnace. The original fluid infrared image is input into a target model, which performs forward inference on the image and outputs the location coordinates, category, and confidence level of the texture features on the molten surface. On the original fluid infrared image, the identified texture features are marked, contoured, and enhanced in grayscale. The enhanced image is then output to obtain a fluid texture enhanced image, which is used for subsequent optical flow calculations.

[0113] In one implementation, training data construction:

[0114] Training data type: Infrared image dataset of molten glass surface;

[0115] Data source: Infrared camera data collected at the glass all-electric melting furnace site;

[0116] Data size: ≥3000 images (80% training set, 10% validation set, 10% test set);

[0117] Label category: texture (unique category for molten surface texture features);

[0118] Labeling format: YOLO format (txt label file, containing category ID, center point coordinates, width, and height);

[0119] Data augmentation: random cropping, contrast adjustment, Gaussian noise, rotation, flipping, brightness perturbation.

[0120] In one implementation, YOLO11 pre-trained weights are loaded, and the first 50% of the backbone network layers are frozen. A pre-constructed infrared texture dataset of molten glass is input, and transfer learning training is initiated. The initial learning rate is set to 1e-4, batch size to 8, input resolution to 640×640, and the number of iterations to 200. The backbone is frozen for the first 100 iterations, and the entire network is unfrozen for fine-tuning in the last 100 iterations. In each iteration, the loss, precision, recall, and mAP@0.5 are calculated. When the validation set mAP does not improve for 10 consecutive iterations, a learning rate decay and early stopping strategy are implemented. The optimal weight file is saved to obtain the target model.

[0121] In one embodiment, a two-dimensional velocity vector map is obtained by calculating the enhanced image using optical flow, including:

[0122] Feature point sets are obtained by extracting features from the enhanced image. The inter-frame displacement dataset is obtained by performing optical flow calculation on the feature coordinate point sets of two temporally adjacent frames. The feature points include: inflection points of liquid surface texture and undulation contour points.

[0123] The feature point velocity vector set is calculated based on the inter-frame displacement dataset. The enhanced image is divided into regions to obtain a uniform computational grid. The feature point velocity vectors in the feature point velocity vector set are added to the uniform computational grid to obtain grid vector data. The grid vector data is padded to obtain complete grid vector data. The complete grid vector data is rendered to obtain a two-dimensional velocity vector map.

[0124] In one implementation, feature points are extracted frame by frame from the enhanced image, and stable feature points such as inflection points of liquid surface texture and undulation contour points are identified and marked, and the feature coordinate point set corresponding to each frame of the image is output.

[0125] In one implementation, two temporally adjacent feature coordinate point sets are selected (the video stream is collected, which consists of consecutive frame images with a fixed frame rate), optical flow constraint calculation is performed, and corresponding feature points in the preceding and following frames are matched according to the basic principle of optical flow. The pixel displacement values ​​of the feature points in the horizontal and vertical directions are calculated to obtain the inter-frame displacement dataset.

[0126] In one implementation, the displacement of feature point pixels is converted into physical flow velocity by combining camera calibration parameters, actual field of view size, and shooting frame rate. The displacement direction corresponds to the flow direction on the surface of the molten glass, and a set of feature point velocity vectors is generated.

[0127] In one implementation, the velocity vector set is subjected to mesh partitioning, the image region is divided into a uniform computational grid, and the single feature point vector is matched to the corresponding grid cell to obtain the grid vector data.

[0128] In one implementation, spatial interpolation is performed on blank grids without feature points, and the flow velocity and direction at that location are calculated based on the surrounding valid vector data to complete the whole-domain vector information and obtain complete grid vector data.

[0129] In one implementation, the complete grid vector data is rendered using vector visualization, with arrows representing the flow direction and arrow length representing the flow velocity, ultimately generating a two-dimensional velocity vector map of the glass melt surface.

[0130] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention should still fall within the scope of the claims of the present invention.

Claims

1. A method for suppressing bubbles in a fully electric melting furnace for borosilicate glass based on three-field synergy, characterized in that, The method includes: The discharge data of the glass furnace is acquired, and the first feature data is obtained by extracting features from the discharge data; Temperature data of a glass melting furnace is acquired, a temperature field is constructed based on the temperature data, and features are extracted from the temperature field to obtain second feature data; the temperature data includes: outer side data and inner side data; High-definition infrared cameras are used to capture fluid images on the surface of glass furnaces. The texture features of the fluid surface are identified using a target model, and these texture features are then added to the fluid images to obtain enhanced images. The enhanced image is calculated using optical flow to obtain a two-dimensional velocity vector map, and the third feature data is obtained by feature extraction from the two-dimensional velocity vector map. The parameter anomaly type is determined based on the first feature data, the second feature data, and the third feature data, and bubble suppression is performed based on the parameter anomaly type.

2. The method for suppressing bubbles in a fully electric melting furnace for borosilicate glass based on three-field synergy as described in claim 1, characterized in that, A temperature field is constructed based on the temperature data, including: The outer data and the inner data are preprocessed to obtain preprocessed outer data and preprocessed inner data, respectively. A three-dimensional model of the glass furnace is obtained, and the three-dimensional model is divided to obtain a spatial mesh model. The preprocessed inner data is added to the spatial grid model to obtain the point assignment grid, and the preprocessed outer data is added to the point assignment grid according to the temperature decay law to obtain the global grid; The empty grid in the global grid is filled with data by spatial interpolation algorithm to obtain complete grid data, and the temperature field is obtained by spatial mapping of the complete grid data.

3. The method for suppressing bubbles in a fully electric melting furnace for borosilicate glass based on three-field synergy as described in claim 1, wherein the training process of the target model before identifying the texture features of the fluid surface through the target model includes: A set of raw fluid infrared images of the surface of a glass furnace is acquired by a high-definition infrared camera. The raw fluid infrared images are then manually labeled to obtain a set of artificially labeled images. The artificially labeled images are then added to the raw fluid infrared images to obtain a training image set. The training image set is input into the target model to obtain the confidence level of the texture features. If the confidence level is greater than the threshold, the target model is deemed to have passed the training.

4. The method for suppressing bubbles in a fully electric melting furnace for borosilicate glass based on three-field synergy as described in claim 1, characterized in that, The target model is an improvement based on the YOLO11 model, including: The target model is obtained by replacing the backbone network with an improved backbone network and the neck structure with an improved neck structure; the YOLO11 model includes a backbone network and a neck structure.

5. The method for suppressing bubbles in a fully electric melting furnace for borosilicate glass based on three-field synergy as described in claim 4, characterized in that, The working process of the improved backbone network includes: Obtain the original feature map A, input the original feature map A into the SRFD module to obtain the first feature map A, input the first feature map A into the CSFPA module, DRFD module and CSFPA module in sequence to obtain the second feature map A, input the second feature map A into the DRFD module and CSFPA module in sequence to obtain the third feature map A, input the third feature map A into the DRFD module, CSFPA module, SPPF module and C2PSA module in sequence to obtain the fourth feature map A, and use the second feature map A, the third feature map A and the fourth feature map A as the output of the improved backbone network.

6. The method for suppressing bubbles in a fully electric melting furnace for borosilicate glass based on three-field synergy as described in claim 5, characterized in that, The process of improving the neck structure includes: The 4th feature map A is upsampled to obtain the 5th feature map A. The 5th feature map A is concatenated with the 3rd feature map A to obtain the 6th feature map A. The 6th feature map A is input into the CSFPA module to obtain the 7th feature map A. The 7th feature map A is upsampled and concatenated with the 2nd feature map A to obtain the 8th feature map A. The 8th feature map A is input into the CSFPA module to obtain the 9th feature map A. The 9th feature map A is input into the DRFD module and then concatenated with the 7th feature map A to obtain the 10th feature map A. The 10th feature map A is input into the CSFPA module to obtain the 11th feature map A. The 11th feature map A is input into the DRFD module and then concatenated with the 4th feature map A to obtain the 12th feature map A. The 12th feature map A is input into the CSFPA module to obtain the 13th feature map A. The 7th feature map A, the 11th feature map A, and the 13th feature map A are used as the output of the improved neck structure.

7. The method for suppressing bubbles in a fully electric melting furnace for borosilicate glass based on three-field synergy according to claim 1, characterized in that, A two-dimensional velocity vector map is obtained by calculating the enhanced image using optical flow, including: Feature point set is obtained by extracting features from the enhanced image, and inter-frame displacement dataset is obtained by performing optical flow calculation on two temporally adjacent feature coordinate point sets; the feature points include: liquid surface texture inflection points and wave contour points; The feature point velocity vector set is calculated based on the inter-frame displacement dataset. The enhanced image is divided into regions to obtain a uniform computational grid. The feature point velocity vectors in the feature point velocity vector set are added to the uniform computational grid to obtain grid vector data. The grid vector data is padded to obtain complete grid vector data. The complete grid vector data is rendered to obtain a two-dimensional velocity vector map.

8. A bubble suppression system for a fully electric melting furnace of borosilicate glass based on three-field synergy, characterized in that, The system includes: The discharge data acquisition module is used to acquire discharge data from the glass furnace and extract features from the discharge data to obtain the first feature data. A temperature data acquisition module is used to acquire temperature data of a glass melting furnace, construct a temperature field based on the temperature data, and extract features from the temperature field to obtain second feature data; the temperature data includes: outer side data and inner side data; The flow field feature acquisition module is used to capture fluid images on the surface of the glass furnace using a high-definition infrared camera, identify the texture features of the fluid surface through a target model, and add the texture features to the fluid image to obtain an enhanced image; The flow field feature extraction module is used to calculate the enhanced image using optical flow to obtain a two-dimensional velocity vector map, and to extract features from the two-dimensional velocity vector map to obtain the third feature data. The bubble suppression module is used to determine the parameter anomaly type based on the first feature data, the second feature data, and the third feature data, and to suppress bubbles based on the parameter anomaly type.

Images