Copper converter blowing end point on-line identification method based on improved target detection model

Abstract

The application relates to the technical field of copper converter blowing process monitoring and automatic control, and discloses a copper converter blowing endpoint online identification method based on an improved target detection model, which comprises the following steps: obtaining real-time image data and process environment parameters in the copper converter blowing process and determining a target identification state; generating a feature extraction signal based on the improved target detection model, the feature extraction signal comprising a melt surface feature, a bubble distribution and a color change index; adopting a self-adaptive optimization algorithm to perform real-time adjustment on the system, effectively correcting global and local identification errors; the system performs real-time monitoring on blowing endpoint related indexes, and adjusts parameters of the model, such as adjusting feature scales, attention weights or iteration frequencies, when deviating from the target state is detected, so as to realize self-adaptive endpoint identification and blowing quality optimization. The application effectively improves the accuracy, real-time performance and blowing process stability of the copper converter blowing endpoint identification.

Description

Technical Field

[0001] This invention relates to the field of copper converter smelting process monitoring and automatic control technology, and more specifically, to an online identification method for the endpoint of copper converter smelting based on an improved target detection model. Background Technology

[0002] Existing online endpoint identification methods for copper converter blowing typically need to address identification errors caused by various factors during dynamic blowing processes, including fluctuations in raw material composition, changes in furnace temperature, unstable oxygen flow, and sensor drift. These errors vary considerably in their manifestations, including global feature misjudgment, local feature extraction deviations, and high-order interference caused by drastic fluctuations in operating conditions. Furthermore, they differ in their causes, degree of impact, and rate of change. Due to the complexity of the causes of identification errors and their dynamic evolution with the blowing process, existing methods have significant limitations in real-time judgment and blowing quality control.

[0003] Currently, most mainstream methods for identifying the end point of refining rely on fixed threshold alarms or static models trained on historical data, typically requiring the design of offline identification schemes using pre-collected sample data. However, this approach is prone to problems such as response lag, insufficient identification accuracy, and poor adaptability in dynamic refining scenarios. When furnace conditions change rapidly or the system operates continuously for extended periods, the identification performance often declines significantly, leading to misjudgments of the end point, over- or under-blowing, and insufficient stability in refining quality, thereby affecting metal recovery rate and product quality. Furthermore, existing identification methods lack the ability to monitor and adjust the refining status in real time during operation, often requiring manual intervention or recalibration to restore system performance.

[0004] In summary, existing technologies are insufficient to effectively address the problems of poor real-time endpoint identification, weak adaptability to complex operating conditions, and insufficient long-term stability in copper converter blowing processes during dynamic processes. In particular, they lack online evaluation and adaptive optimization capabilities for the blowing state. Therefore, there is an urgent need to propose a new online endpoint identification and blowing optimization method to improve the performance of copper converter blowing in terms of dynamic error suppression, real-time response, and blowing consistency, thereby meeting the requirements of high-quality smelting and fully automated process control for accurate endpoint judgment and process stability. Summary of the Invention

[0005] In view of this, the present invention proposes an online identification method for the end point of copper converter blowing based on an improved target detection model, aiming to solve the problems of poor real-time performance of end point judgment, weak adaptability to complex working conditions, insufficient long-term operational stability, and lack of online evaluation and adaptive optimization capabilities for blowing status in existing identification methods during dynamic blowing processes.

[0006] This invention proposes an online identification method for the endpoint of copper converter blowing based on an improved target detection model, including: Real-time image data and process environment parameters are acquired during the copper converter blowing process, and the target recognition status is determined based on the image data and process environment parameters. The process environment parameters include blowing temperature, oxygen flow rate, slag composition and melt stirring speed. An improved target detection model is constructed based on the target recognition status, and corresponding feature extraction signals are generated according to the improved target detection model. The feature extraction signals include melt surface features, bubble distribution, and color change indicators. An adaptive optimization algorithm is used to adjust the copper converter blowing system in real time to effectively correct identification errors; In the actual blowing process, the improved target detection model is input based on the feature extraction signal, and relevant indicators at the blowing endpoint are monitored in real time; When the blowing endpoint index is detected to deviate from the target identification state, at least one parameter of the improved target detection model is adjusted in a timely manner to achieve reliable endpoint identification and further optimize the overall blowing quality.

[0007] Furthermore, the acquisition of process environment parameters includes: Acquire environmental vibration intensity, furnace pressure changes, and external disturbance parameters; The threshold parameters in the improved target detection model are corrected based on changes in environmental vibration and pressure.

[0008] Furthermore, the generation of the feature extraction signal includes: In the initial stage of recognition, a gradual multi-scale feature fusion is adopted. The fusion scale change amplitude and the signal maintenance time after each fusion are reasonably configured to enable the melt features to be smoothly extracted to the target accuracy and effectively suppress noise interference. Once the recognition quality reaches the target recognition state and enters a stable phase, it switches to a maintenance mode with low computational complexity.

[0009] Furthermore, the adaptive optimization algorithm includes: The first optimization layer, located at the front end of the model, is used to absorb global recognition errors; A high-precision correction layer is applied over the first optimization layer to dynamically adjust local feature deviations; An integrated computing framework located at the system backend provides overall support and reliably interfaces with the refining controller.

[0010] Furthermore, the adaptive optimization algorithm also incorporates an auxiliary optimization mechanism, including: The attention module and multimodal fusion strategy are implemented, and the weights of the optimization mechanism are controlled within a range below a preset weight threshold.

[0011] Furthermore, the real-time monitoring of indicators related to the refining endpoint includes: Obtain the actual distribution information of the recognition error; And calculate the offset of the current endpoint judgment based on the distribution information.

[0012] Furthermore, at least one parameter of the adjusted and improved target detection model includes: The scale parameter of the feature extraction signal is adjusted, wherein the scale adjustment amount is determined based on the deviation between the real-time monitored refining endpoint index and the target recognition state. Adjust the attention weights of the improved object detection model to minimize errors while maintaining stable recognition. The iteration frequency of the improved target detection model was adjusted to appropriately extend the optimization time and reduce the overall computational cost while meeting the response speed requirements.

[0013] Furthermore, the endpoint identification process also includes: After the blowing quality reaches a stable state, switch to compensation recognition mode; The computational intensity of the compensation identification is lower than the initial adjustment intensity set in the improved target detection model.

[0014] Furthermore, it also includes: The aging degree of the system is assessed based on the cumulative blowing time and the number of adjustments. And when the aging degree exceeds the preset threshold, the safety optimization range in the improved target detection model is adjusted accordingly.

[0015] Furthermore, the improved target detection model is a neural network model trained based on historical blowing quality data and actual adjustment response data; The neural network model can adaptively optimize the combination of parameters of the feature extraction signal based on the deviation between the real-time monitored recognition state and the target state.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: By acquiring real-time image data and process environment parameters of the copper converter blowing process, and determining the target recognition status based on this, the system can systematically identify key factors affecting endpoint judgment. The resulting benefits include significantly improved characterization capabilities for complex and dynamically changing blowing conditions, enabling the system to comprehensively consider the inherent characteristics of the melt and external interference, thus establishing a reliable benchmark for accurate recognition, reducing reliance on fixed thresholds or offline models, and improving the comprehensiveness and accuracy of endpoint analysis. Secondly, an improved target detection model is constructed based on the target recognition status, and feature extraction signals containing parameters such as multi-scale features and attention weights are generated according to the model, achieving real-time and intelligent analysis of the melt surface state and dynamic features. The advantages of this technology lie in its ability to fully utilize the powerful representational capabilities of deep learning, achieving smooth information extraction and noise suppression through gradual feature fusion, effectively avoiding model oscillations in the initial recognition stage, and maintaining efficient operation with low computational complexity after steady-state operation, thereby improving the stability and efficiency of the recognition process. Thirdly, an adaptive optimization algorithm with multi-layer optimization mechanisms is used to adjust the system in real time, directly absorbing and offsetting local and global recognition errors at the algorithm level. Its beneficial effects are manifested in that the system can simultaneously correct global misjudgments and local deviations, and enhance key features by introducing mechanisms such as attention modules, thereby maintaining high recognition accuracy and judgment consistency even in complex interference scenarios. Finally, by monitoring relevant indicators of the blowing endpoint in real time and adaptively adjusting the model parameters when deviations from the target state are detected, the system can automatically complete the feedback control closed loop. The beneficial effect of this mechanism is that it enables the entire endpoint identification process to have online self-optimization capabilities, and can adjust the identification strategy and safety optimization range in a timely manner when operating conditions change, sensor performance drifts, or during long-term continuous operation, thereby ensuring the long-term stability, operating condition adaptability, and overall reliability of the blowing endpoint judgment. Attached Figure Description

[0017] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart illustrating the online identification method for the end point of copper converter blowing based on an improved target detection model, provided in an embodiment of the present invention. Figure 2 This is a flowchart illustrating the online identification method for the end point of copper converter blowing based on an improved target detection model, provided in an embodiment of the present invention. Detailed Implementation

[0018] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] like Figures 1-2 As shown in some embodiments of this application, this embodiment provides an online identification method for the endpoint of copper converter blowing based on an improved target detection model, including: Step S100: Obtain real-time image data and process environment parameters during the copper converter blowing process, and determine the target recognition status.

[0020] Specifically, step S100 includes: acquiring real-time multispectral / visible light image data of the copper converter blowing process and key process environment parameters through an industrial camera, infrared thermal imager, furnace mouth visual sensing unit, process database stored internally in the equipment, or distributed control system interface. These process environment parameters mainly include the blowing temperature (obtained via thermocouples or infrared thermometers, typically ranging from 1150℃ to 1250℃) and oxygen flow rate (unit: Nm³). 3 The system dynamically determines the target recognition status of the current blowing stage based on image features and process parameters, including the following parameters: slag composition (Fe / SiO2 ratio, CaO content, etc. calculated based on online XRF analysis or historical charging models), melt stirring speed (indirectly reflected by bottom blowing gas flow rate or converter tilting rate), blowing air pressure, and cycle time. This target status includes preset endpoint judgment feature thresholds (e.g., specific color areas on the melt surface account for more than 85%, bubble size and distribution density reach a stable range, flame brightness / color feature vector is located at the target cluster center), target recognition accuracy requirements (e.g., not less than 98.5%), overall misjudgment rate tolerance (e.g., less than 1%), and recognition response delay upper limit (e.g., less than 3 seconds).

[0021] In specific implementation, preferably, the system first reads the basic process parameters of the current furnace run from the metallurgical plant's manufacturing execution system or process knowledge base, such as copper matte grade (45% to 65%), target blister copper grade (≥98.5%), planned blowing time, and flux addition ratio. Simultaneously, the system acquires high-frame-rate image data (visible and near-infrared bands, frame rate ≥25fps) in real time using multi-source vision sensors deployed at the furnace opening, and obtains environmental parameters provided by the DCS system in real time via the industrial bus, including but not limited to real-time blowing temperature (fluctuation within ±20℃), instantaneous and cumulative oxygen flow rates, furnace pressure changes (±5kPa), environmental vibration intensity (caused by adjacent equipment), and external interference parameters caused by raw material composition fluctuations or operational adjustments. Furthermore, based on the influence of temperature on melt fluidity and surface oxidation state, the correlation between oxygen flow rate and reaction kinetics and bubble formation characteristics, and the interference of external vibration on image stability, the system dynamically corrects the key feature extraction thresholds and classification confidence parameters in the improved target detection model in real time. For example, when the detected temperature is below 1180℃, the system lowers the threshold for identifying the dark red area representing Fe3O4 formation by 5%-10% in color recognition to more sensitively capture the initial signs of "black copper." When the oxygen flow fluctuates drastically, the system increases the weighting of the stability of the bubble size sequence. This ensures that the target recognition state is always highly adapted to the current actual blowing conditions, avoiding recognition deviations caused by condition drift.

[0022] Understandably, compared to traditional methods that rely solely on a single temperature point or fixed time threshold for endpoint determination, or that use only static image template matching, this step, by synchronizing multimodal image data and full-process parameters in real time and dynamically determining the target identification state, can more comprehensively and accurately reflect the essential state evolution of copper converter blowing in the actual dynamic reaction process. For example, when the grade of copper matte entering the furnace is too high, potentially extending the blowing time, or when fluctuations in the flux SiO2 content affect slag properties and thus alter the reflective characteristics of the melt surface, this method can directly reflect these changes through real-time parameter fusion and dynamic adjustment of the target state. This lays a reliable data fusion foundation for subsequent feature focusing and intelligent judgment in the model, significantly improving the robustness of endpoint identification and its adaptability to complex operating conditions. Simultaneously, this step effectively reduces the absolute reliance on the operator's personal experience, enabling the system to quickly adapt to different raw material batches and process fine-tuning, significantly improving deployment efficiency and stability in actual large-scale continuous production.

[0023] In a specific embodiment of this application, the above steps are implemented as follows: For the blowing process of a PS converter (Φ4.0m×9.0m) in a large copper smelter, before the start of a single blowing cycle, the system's main control module first automatically reads the planned parameters for that cycle from the MES: 150t of copper matte to be processed, 58% copper matte grade, 98.8% target blister copper grade, and a planned blowing time of 125 minutes. Simultaneously, a high-definition high-temperature resistant camera (equipped with a narrow-band filter) and an infrared thermal imager at the furnace mouth are activated to collect real-time images of the melt surface and temperature field distribution. The DCS real-time data interface inputs the current parameters: initial temperature 1220℃, oxygen flow rate set to 18000Nm³. 3 The system operates at a constant temperature of +2 kPa within the furnace, with an environmental vibration monitoring value of 0.5 g. Based on the initial high-temperature conditions, the system initially sets the confidence threshold for determining the key characteristic in image recognition marking the end of the slag-forming period and the transition to the copper-forming period—the transformation of the melt surface from a chaotic, foamy state to a smooth, metallic luster accompanied by the appearance of specific red-yellow areas—to 0.92. Simultaneously, based on the high initial oxygen flow rate, the system increases the sampling frequency of the bubble feature analysis module to 10 Hz to capture rapidly changing dynamic characteristics. Through the process dashboard, operators can visually observe the target recognition status generated after the system integrates multi-source information: the overall confidence level for the endpoint determination must be >0.95, the proportion of the bright red main color area in the melt must be >82%, the average bubble size in five consecutive frames must be stable within the range of 3 mm ± 0.5 mm, and the output delay of the recognition result must be <2.5 seconds. This provides a highly personalized benchmark reference tightly coupled with real-time operating conditions for subsequent accurate model inference.

[0024] The above scenarios are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0025] Step S200: Construct an improved target detection model based on the target recognition status, and generate corresponding feature extraction and recognition signals.

[0026] Specifically, step S200 includes: constructing a dedicated improved target detection model based on the target recognition state determined in step S100. This model can be a deep learning model trained on a large amount of historical blowing video data, successful / failure endpoint sample images, process parameter time series data, and metallurgical reaction mechanism knowledge (such as an improved YOLO series, a hybrid architecture of Faster R-CNN and Transformer), or it can be an experience-data hybrid driven model that combines computational fluid dynamics simulation results with image features. Subsequently, the model is used to process the real-time input multi-frame image sequence to generate multi-level feature extraction signals and preliminary recognition decision signals, mainly including macroscopic color and texture feature vectors of the melt surface, size / distribution / evolution dynamic features of the bubble group, brightness / chromaticity indices of specific areas (such as the root of the furnace flame, the specular reflection area of ​​the melt), and a preliminary endpoint arrival probability value.

[0027] In its implementation, the system preferentially employs a multi-scale feature pyramid network with embedded channel and spatial dual attention mechanisms as the core target detection model. This model is pre-trained offline using over 5000 annotated video clips accumulated over the past few years (annotation information includes the endpoint time and typical features of each stage). During online operation, it undergoes incremental learning and fine-tuning through continuously collected new data and operator-confirmed endpoint results. The model can extract and fuse features at different levels of abstraction with high precision: shallow networks capture details such as edges and textures (e.g., the edges of small bubbles, cracks in the slag shell); deep networks understand semantic information (e.g., large areas of metallic luster, churning slag phase foam). At the critical end of the copper-making stage of the blowing process (i.e., the endpoint recognition stage), the model first initiates a "progressive feature focusing" strategy: in the initial recognition stage, the model simultaneously focuses on global color distribution and local dynamic features, but after each preset time window (e.g., 30 seconds), the system adaptively adjusts the weights of different feature channels based on the difference between the current recognition confidence and the target state. For example, when color features tend to stabilize but bubble features are still changing, the system gradually reduces the weight of color features while increasing the depth of analysis on the autocorrelation of bubble size sequences. This weight change is designed to be smooth and gradual (each adjustment does not exceed 15% of the total weight), and is maintained for a certain period of time (e.g., 10-20 seconds) after each adjustment to observe the effect, ensuring that the model's attention can smoothly and directionally shift to the most significant changing features, effectively suppressing misjudgment oscillations caused by sensor noise or brief fluctuations in operating conditions. When the endpoint probability value output by the model is continuously higher than the threshold set for the target recognition state (e.g., 0.93) and reaches the preset stable duration (e.g., 1 minute continuously), the system determines that it has entered the "recognition maintenance phase," at which point the model will switch to a computationally optimized lightweight inference mode. In this mode, the model no longer performs dense feature extraction at the full scale, but only performs high-frequency (e.g., once per second) rapid verification on a few key feature channels (such as pre-determined metallic ROI regions), and reduces the complete inference frequency of the complex network to once every 10-15 seconds. This significantly reduces the power consumption and computing resource consumption of edge computing devices while ensuring that the endpoint monitoring is not lost, thus providing a guarantee for long-term continuous production.

[0028] Understandably, the model operation strategy combining progressive feature focusing with lightweight steady-state maintenance can fully leverage the powerful feature representation and pattern recognition capabilities of deep learning models, while minimizing false alarms (mistaking normal process fluctuations for endpoints) or missed alarms (failing to identify endpoints in a timely manner) that are easily caused by traditional single-threshold methods or fixed-cycle scanning methods. On the one hand, progressive attention shifting simulates the observation logic of experienced operators, focusing on different priorities at different stages, making the recognition process more robust. On the other hand, the steady-state maintenance mode significantly reduces the average computational load of the system (by 40% to 60%), reduces hardware heat generation and energy consumption, and also reduces the instability risks that may arise from prolonged high-load operation of the software system, providing crucial support for the reliable 24 / 7 operation of the model in harsh industrial environments. Furthermore, the model's online incremental learning capability enables it to continuously adapt to new feature patterns brought about by changes in raw material sources and process route adjustments, achieving continuous evolution of recognition capabilities.

[0029] In a specific embodiment of this application, the above steps are implemented as follows: For the aforementioned PS converter blowing process, after setting the target state in step S100, the system loads a pre-trained improved YOLO-Transformer hybrid model. When the blowing process reaches approximately 110 minutes (predicting the endpoint sensitive window based on historical data), the model initiates progressive focusing. In the initial stage, the model outputs in parallel: color feature confidence score 0.88 (based on the melt body being bright red-yellow), and bubble stability feature confidence score 0.75 (bubble size still fluctuates slightly). Since the color feature is close to the target but not completely stable, the model fine-tunes the color feature weight from 0.5 to 0.55 and the bubble feature weight from 0.5 to 0.45 within the first 30-second window. Subsequently, it is observed that the bubble size variation coefficient begins to decrease significantly. In the second window, the model reverts the bubble feature weight back to 0.48 and introduces a rapid analysis of the flame color spectrum at the furnace opening as an auxiliary feature (initial weight 0.02). By the 115th minute, the model's overall confidence score reaches 0.94 and remains there for 45 seconds. The system then triggered a switch to lightweight maintenance mode: color histogram matching is performed only on the ROI region at the center of the melt every second (with minimal computational cost), and a full model inference is initiated every 10 seconds for verification. The full model quickly outputs results: endpoint probability of 0.96, mainly based on the color region accounting for 85.2%, and the bubble dynamic index entering a stable range. Throughout the near-endpoint identification process, the model's attention shifted smoothly, and the computational resource usage showed a healthy curve of "rapid decay after peak," successfully and accurately capturing the endpoint signal near the planned blowing time without abnormal oscillations.

[0030] Similarly, the above scenarios are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0031] Step S300: Use an adaptive optimization algorithm to adjust the recognition system in real time to correct recognition errors.

[0032] Specifically, step S300 includes: real-time post-processing and calibration of the preliminary identification results based on an adaptive optimization algorithm suite that works in conjunction with an improved target detection model. This algorithm suite constitutes a dynamic error correction system at the software level. It mainly includes a first-level global optimization layer at the front end of data processing (primarily responsible for macroscopic rationality verification and gross error absorption of the identification results based on the process mechanism model), a second-level high-precision correction layer covering the output of the identification model (dedicated to dynamic fine-tuning and smoothing filtering of the probability values ​​or bounding boxes of the model output based on real-time temporal context information), and a third-level integrated decision framework at the decision output end (providing the final decision logic based on multi-criteria fusion and reliably interfaced with the upper-level blowing control system). Simultaneously, various auxiliary optimization mechanisms are introduced into the optimization process as "software compensation media," such as a time-series predictor based on Kalman filtering or particle filtering, a metallurgical rule inference engine based on knowledge graphs, and a multi-model voting integration strategy.

[0033] In practical implementation, the first-level global optimization layer incorporates a simplified material and heat balance model for the blowing process. Based on real-time cumulative oxygen consumption and flux addition data, it can estimate the theoretical position of the current blowing process. When the endpoint signal output by the target detection model deviates from the theoretical process beyond a safe range (e.g., reporting a high-probability endpoint just halfway through the blowing time), this layer triggers an alarm and significantly reduces the confidence weight of the model output, or directly marks it as a suspicious result, thus absorbing global misjudgments caused by temporary image sensor obstruction or furnace splash interference. The second-level high-precision correction layer focuses on handling local, high-frequency identification fluctuations. It uses a sliding time window (e.g., 30 seconds to 2 minutes) to analyze the probability value sequence continuously output by the model, applying dynamic time warping algorithms or adaptive low-pass filters to identify and suppress short-duration probability "glitch" that does not conform to the inertia of process changes. Simultaneously, it responds quickly to real, continuous trend changes, achieving temporal smoothing and precise fine-tuning of the identification results. The third-level integrated decision framework receives signals optimized by the first two levels and integrates the outputs of various auxiliary optimization mechanisms (such as the future trend given by the time series predictor and the rationality judgment given by the rule engine based on slag alkalinity). Following a preset fusion strategy (such as weighted average and DS evidence theory), it generates the final endpoint determination instruction with a credibility rating. This adaptive optimization algorithm system, based on the initial identification by the target detection model, further absorbs and offsets residual identification errors, sudden interference, and the uncertainty of the model itself from multiple software levels, including data consistency, time series smoothness, and process rationality, achieving more robust and reliable endpoint determination.

[0034] Understandably, the organic integration of the adaptive optimization algorithm system and the front-end target detection model forms a two-layer intelligent decision-making mechanism that coordinates "perception and optimization," maintaining accuracy in judgments across a wider range of operating conditions and more complex interference scenarios. On one hand, the target detection model acts as a high-performance, high-resolution "sensor," responsible for extracting subtle features from complex images; on the other hand, the optimization algorithm system acts as a "brain" with process knowledge, responsible for cross-validating, denoising, and logically reasoning the sensory information, thereby significantly improving the overall system's intelligence level and industrial usability. Furthermore, this system is entirely software-based, possessing high flexibility and configurability, facilitating customized adjustments based on the specific conditions and requirements of different factories.

[0035] In a specific embodiment of this application, the above steps are implemented as follows: In step S200, at a certain moment, due to interference from dense smoke and dust briefly emanating from the furnace opening, the model outputs a peak in endpoint probability lasting approximately 8 seconds (jumping from 0.85 to 0.91). The sliding window filter (window length 45 seconds) of the second-level high-precision correction layer immediately identifies that this peak is inconsistent with the trend of its preceding and following data and that its duration is too short, quickly smoothing it and correcting the output probability at that moment back to approximately 0.86. Almost simultaneously, the theoretical blowing progress calculated by the first-level global optimization layer based on real-time cumulative oxygen consumption is only 82%, far from reaching the endpoint range, thus generating a "strongly unreasonable" flag. The third-level integrated decision framework receives the model's original probability (already smoothed by the second level), the smoothed probability of the second level, the "strongly unreasonable" flag from the first level, and the "smoke and dust interference is common" prompt from the rule engine. The decision framework ultimately determines that the result at that moment is "not the endpoint," with a comprehensive confidence level of only 0.15, and notifies the monitoring module to strengthen the inspection of image quality. The entire process is completed within seconds, effectively avoiding a false alarm caused by momentary interference and ensuring the continuous and stable operation of the blowing process.

[0036] The above scenarios are merely preferred embodiments of the present invention and are not intended to limit the present invention.

[0037] Step S400: Monitor relevant indicators at the end of the blowing process in real time, and adjust model parameters when deviations are detected to achieve adaptive identification and blowing optimization.

[0038] Specifically, step S400 includes: throughout the entire actual refining process, the system uses an independent performance monitoring and evaluation module to track multiple indicators directly or indirectly related to the quality of endpoint identification in real time. These indicators include not only the endpoint probability value and feature confidence directly output by the model, but also process consistency indicators (such as the variance of identification results for N consecutive frames), data quality indicators (such as image clarity, contrast, and occlusion alarms), and process correlation indicators (such as the degree of consistency between the identification results and the trend of oxygen consumption rate changes). When the monitoring module detects that any one or more indicators deviate from the target identification state set in step S100 by more than a preset threshold, the system immediately triggers a parameter adaptive adjustment process. This process is not a simple alarm, but rather an automatic online, iterative optimization and fine-tuning of the key adjustable parameters of the improved target detection model and / or adaptive optimization algorithm. The adjustment objects include, but are not limited to: the scale parameters of feature fusion within the model (adjusting their contribution weights based on which scale feature is currently the most unstable), the spatial / channel attention weight distribution in the attention mechanism, the window length or cutoff frequency of the filter in the second-level correction layer, and the activation strategy and weights of the auxiliary optimization mechanism. The goal of the adjustment is to bring the system output back to the target state in the shortest possible time with minimal computational overhead. In addition, the system accumulates data such as the total number of blowing furnace cycles, the number of model parameter adjustment triggers, and the statistical deviation between historical identification results and manual confirmation results to comprehensively evaluate the degree of "performance drift" or "scenario adaptability aging" of the model or algorithm. When the evaluated aging degree exceeds a preset threshold (e.g., an increase of 20% in the average number of adjustments over 10 consecutive furnace cycles, or a continuous increase in the deviation from manual confirmation), the system will automatically initiate a more in-depth safety optimization range recalibration process, such as widening or tightening the adjustment range of certain parameters, or even prompting for offline model retraining data collection.

[0039] In the implementation, the monitoring module performs metric calculations at a frequency synchronized with or higher than the image acquisition frame rate (e.g., 1Hz). If a deterioration in the "process consistency metric" is detected (e.g., the endpoint probability oscillates drastically by more than 0.15 within one minute), the system diagnostic module first analyzes whether the problem is a data quality issue (e.g., image blurring) or a model adaptability issue. If data quality is ruled out, parameter adjustment is initiated: assuming the diagnostics find a sharp increase in the output variance of the color feature channel while the texture feature remains stable, the adaptive algorithm temporarily reduces the fusion weight of the color feature (by 10%) and correspondingly increases the weight of the texture feature, while possibly slightly increasing the filtering intensity of the second-level correction layer to quickly stabilize the output. All adjustments are performed within a constrained parameter space to ensure that the system does not behave abnormally due to a single adjustment. Simultaneously, the effect of each adjustment is recorded and used for training the aging evaluation model.

[0040] Understandably, this real-time monitoring and adaptive parameter adjustment mechanism transforms the entire endpoint identification system from a static, fixed tool after deployment into a "living" system with online self-learning, self-diagnosis, and self-optimization capabilities. It effectively addresses long-term, slow systemic drift caused by seasonal changes in plant lighting conditions, slow changes in furnace mouth shape due to furnace lining erosion, gradual changes in image quality due to camera lens contamination, and new visual features introduced by changes in raw material sources. This ensures long-term consistency in identification performance throughout the system's entire lifecycle (potentially several years), significantly reducing maintenance costs and process risks caused by performance degradation. This mechanism is one of the core components for achieving truly "unmanned" intelligent blowing.

[0041] In a specific embodiment of this application, the above steps are implemented as follows: After the PS converter in the aforementioned smelter has been running continuously for two weeks, the monitoring module detected a trend: during the middle of each day's shift (10:00 AM to 11:00 AM), due to strong reflections caused by direct sunlight from the roof windows of the plant onto the furnace opening area, the model's extraction of the specular gloss area of ​​the melt showed slight instability for several consecutive days, resulting in a slight increase in the fluctuation of the endpoint probability near the critical value (0.90 to 0.93). Although no false alarms have been triggered, the "process consistency index" has been consistently lower than the target value. When the system detects this phenomenon again at 10:15 AM on the third day, it automatically triggers parameter adjustment. The diagnostic module analyzes that the reflection mainly affects the highlight part of the image, thereby affecting the high brightness range of the color histogram. The adjustment strategy is as follows: 1. Temporarily lower the upper threshold of the V (luminance) channel in the HSV space during color feature extraction by 5% to weaken the impact of highlight areas; 2. Temporarily and slightly reduce the attention weight of common highlight areas in the central part of the image in the model's attention mechanism; 3. Slightly increase the time window length of the second-level corrector from 50 seconds to 55 seconds to enhance smoothing ability. Approximately 5 minutes after the adjustment was implemented, monitoring indicators showed that consistency had returned to normal levels. Simultaneously, the system recorded the reason for this adjustment as "periodic lighting interference" and stored this pattern in association with time. When the evaluation model determines that this type of adjustment caused by ambient light at a fixed time has occurred regularly more than 5 times, the system automatically performs a permanent fine-tuning of the safe optimization range of relevant parameters within that time period (10:00-11:30) (e.g., allowing the V channel threshold to have a wider adjustment range during this time period), thereby achieving persistent adaptive capability against interference during specific time periods, eliminating the need for frequent subsequent adjustments.

[0042] The above scenarios fully demonstrate that by combining continuous indicator monitoring with intelligent parameter adaptive adjustment, this method enables the system to proactively adapt to complex and dynamic industrial environments, significantly improving the long-term stability and generalization ability of overall recognition performance.

[0043] Step S500: Deploy the optimized identification decisions and parameter configurations to the blowing control system to achieve closed-loop intelligent blowing optimization.

[0044] Specifically, step S500 includes: based on the final and stable endpoint identification decision ("endpoint reached" or "endpoint not reached", with high confidence) obtained through iterative optimization in step S400, and the current optimal model and algorithm parameter configuration determined after adaptive adjustment, it is completely and reliably sent to the core control system (such as DCS or PLC) of the copper-to-carbon refining process. In this closed loop, the main refining control module acts as the execution center, responsible for receiving endpoint instructions from the identification system, real-time process data, and manual instructions from the interactive interface. It parses the endpoint instructions into specific equipment control sequences, such as: sending instructions to stop the oxidation blower and close the oxygen lance valve, starting the converter to tilt to the copper tapping position, and triggering the feeding system to prepare for the next batch of cold charge, etc. At the same time, the main module packages and archives the full cycle data of this refining process (including images, intermediate identification results, adjustment records, and final process results) for model iterative updates. The entire system tightly couples the image acquisition and processing unit, process data interface unit, intelligent identification and optimization calculation unit (running S100-S400), blowing control execution unit, and human-machine interaction and monitoring unit through high-speed industrial Ethernet or real-time bus, realizing a complete closed-loop intelligent blowing optimization process from "multi-source information perception - intelligent status recognition - online self-optimization - precise control execution - knowledge accumulation".

[0045] In its implementation, upon receiving a high-confidence (e.g., >0.97) "endpoint reached" signal, the main blowing control module initiates a standardized operational sequence that includes a brief confirmation delay (e.g., 2-3 seconds to prevent single-frame misjudgment) and safety interlock checks (e.g., ensuring the oxygen lance has been raised to a safe height) before issuing critical control commands to the underlying equipment. Simultaneously, it associates and binds data such as the "actual endpoint time," "recommended endpoint time by the identification system," and "snapshots of key visual and process parameters" of this blowing operation with the production order and quality test results (after furnace exit) in the MES, forming a complete digital archive for each furnace. The human-machine interface provides operators with a panoramic visualization interface, displaying not only real-time identification results and confidence levels, various monitoring indicator curves, and heatmaps of model focus, but also interfaces for "manual confirmation," "manual veto," and "parameter intervention," ensuring ultimate human-in-the-loop control.

[0046] Understandably, through the aforementioned modular and closed-loop system integration design, this method achieves a complete intelligent closed loop from "perception-cognition-decision-execution-learning." Compared to traditional manual operation relying on visual inspection or single instrument judgment, this intelligent closed-loop system can not only complete high-precision status identification and judgment within milliseconds / seconds, but also seamlessly, automatically, and safely convert the judgment results into production equipment actions, continuously optimizing its own identification capabilities in the process. This greatly improves the real-time performance, accuracy, consistency, and traceability of the blowing endpoint control, reduces fluctuations caused by human factors, and has revolutionary significance for improving crude copper quality, reducing energy and material consumption, extending furnace lining life, and achieving standardized and efficient production. It is particularly suitable for modern, large-scale copper smelting production lines that pursue ultimate efficiency and stability.

[0047] In a specific embodiment of this application, the above steps are implemented as follows: For the aforementioned PS converter, after completing the monitoring and adaptive adjustment in step S400, the system outputs a final judgment at 122 minutes and 35 seconds into the blowing process: the endpoint has been reached, with a comprehensive confidence level of 0.983. After receiving the signal, the main blowing control module first checks the oxygen lance position (already in the lifting position) and image quality (good). After a 2-second delay, the confidence level remains above 0.980. The main module then executes the following in sequence: 1. Sending a closing command to the oxygen regulating valve; 2. Sending a frequency reduction and shutdown command to the oxidation blower; 3. Sending a command to the converter tilting control system to "rotate to the copper tapping position (angle X)"; 4. Displaying a large-font prompt on the operator HMI: "The blowing endpoint has been reached, copper tapping process is being executed," and recording all action timestamps. The entire control sequence is completed automatically within 5 seconds, and the blowing process ends smoothly. Subsequently, a digital file for this furnace run was automatically generated, which recorded that the system identified the finish line 2 minutes and 25 seconds ahead of schedule, and based on historical data, it was estimated that approximately 450 Nm³ of oxygen could be saved. 3 The operator clicks "confirm" on the interface, and the file is synchronized to the MES and the model incremental learning database. In the following month of continuous production, the system automatically determined the endpoint for 156 heats, with a manual verification accuracy rate of 99.4%. The average blowing time fluctuation per heat was reduced by 35% compared to the manual control period, and the stability of the crude copper grade was significantly improved.

[0048] In another preferred embodiment, the present invention also provides an online identification and optimization system for the intelligent blowing endpoint of copper converter using the above-described method. This system includes a multi-source data acquisition module (industrial camera, infrared thermal imager, process data interface), an edge computing and intelligent processing unit (built-in improved target detection model, adaptive optimization algorithm, monitoring and evaluation module), a blowing control execution module (integrated into DCS / PLC), a data storage and knowledge management module, and a human-machine collaborative interaction module. The network connections, data flows, and control flows between these modules completely correspond to the detailed descriptions in steps S100 to S500 of the aforementioned method. By organically and deeply integrating industrial visual perception, deep learning intelligent recognition, online adaptive optimization, precise closed-loop control, and metallurgical knowledge accumulation, this system achieves intelligent determination and fully automatic optimization control of the copper converter blowing endpoint under various complex raw materials and operating conditions. This invention demonstrates significant practical value, economic benefits, and broad prospects for industrialization in multiple aspects, including improving metal recovery rates, saving energy and reducing consumption, stabilizing product quality, reducing worker labor intensity, and realizing the digital and intelligent transformation of smelting.

[0049] It is understood that the online identification method and system for the copper converter blowing endpoint based on the improved target detection model in the above embodiments have the same or highly similar technical advantages and beneficial effects. The main differences between the aforementioned embodiments are reflected in the specific process parameter range, visual sensor selection, deep learning model architecture details, optimization algorithm combination, and interface protocol of the factory control system. Reasonable adjustments, functional equivalent substitutions, or detailed improvements made by those skilled in the art without departing from the core spirit and substantive technical solution of this invention should be considered to fall within the scope of protection of the claims of this invention.

[0050] It is understood that the online identification method for the end point of copper converter blowing based on the improved target detection model in the above embodiments has the same beneficial effect, and will not be described in detail here.

[0051] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program goods. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program goods embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0052] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program goods according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0053] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0054] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0055] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for online identification of the endpoint of copper converter blowing based on an improved target detection model, characterized in that, Includes the following steps: Real-time image data and process environment parameters are acquired during the copper converter blowing process, and the target recognition status is determined based on the image data and process environment parameters. The process environment parameters include blowing temperature, oxygen flow rate, slag composition and melt stirring speed. An improved target detection model is constructed based on the target recognition status, and corresponding feature extraction signals are generated according to the improved target detection model. The feature extraction signals include melt surface features, bubble distribution, and color change indicators. An adaptive optimization algorithm is used to adjust the copper converter blowing system in real time to effectively correct identification errors; In the actual blowing process, the improved target detection model is input based on the feature extraction signal, and relevant indicators at the blowing endpoint are monitored in real time; When the blowing endpoint index is detected to deviate from the target identification state, at least one parameter of the improved target detection model is adjusted in a timely manner to achieve reliable endpoint identification and further optimize the overall blowing quality.

2. The online identification method for the end point of copper converter blowing based on an improved target detection model as described in claim 1, characterized in that, The process environment parameters to be acquired include: Acquire environmental vibration intensity, furnace pressure changes, and external disturbance parameters; The threshold parameters in the improved target detection model are corrected based on changes in environmental vibration and pressure.

3. The online identification method for the end point of copper converter blowing based on an improved target detection model as described in claim 1, characterized in that, The generation of the feature extraction signal includes: In the initial stage of recognition, a gradual multi-scale feature fusion is adopted. The fusion scale change amplitude and the signal maintenance time after each fusion are reasonably configured to enable the melt features to be smoothly extracted to the target accuracy and effectively suppress noise interference. Once the recognition quality reaches the target recognition state and enters a stable phase, it switches to a maintenance mode with low computational complexity.

4. The online identification method for the end point of copper converter blowing based on an improved target detection model as described in claim 1, characterized in that, The adaptive optimization algorithm includes: The first optimization layer, located at the front end of the model, is used to absorb global recognition errors; A high-precision correction layer is applied over the first optimization layer to dynamically adjust local feature deviations; An integrated computing framework located at the system backend provides overall support and reliably interfaces with the refining controller.

5. The online identification method for the end point of copper converter blowing based on an improved target detection model as described in claim 1, characterized in that, The adaptive optimization algorithm also incorporates an auxiliary optimization mechanism, including: The attention module and multimodal fusion strategy are implemented, and the weights of the optimization mechanism are controlled within a range below a preset weight threshold.

6. The online identification method for the end point of copper converter blowing based on an improved target detection model as described in claim 1, characterized in that, The real-time monitoring indicators related to the end point of the blowing process include: Obtain the actual distribution information of the recognition error; And calculate the offset of the current endpoint judgment based on the distribution information.

7. The online identification method for the end point of copper converter blowing based on an improved target detection model as described in claim 1, characterized in that, At least one parameter of the adjusted and improved target detection model includes: The scale parameter of the feature extraction signal is adjusted, wherein the scale adjustment amount is determined based on the deviation between the real-time monitored refining endpoint index and the target recognition state. Adjust the attention weights of the improved object detection model to minimize errors while maintaining stable recognition. The iteration frequency of the improved target detection model was adjusted to appropriately extend the optimization time and reduce the overall computational cost while meeting the response speed requirements.

8. The online identification method for the end point of copper converter blowing based on an improved target detection model as described in claim 1, characterized in that, The endpoint identification process also includes: After the blowing quality reaches a stable state, switch to compensation recognition mode; The computational intensity of the compensation identification is lower than the initial adjustment intensity set in the improved target detection model.

9. The online identification method for the end point of copper converter blowing based on an improved target detection model as described in claim 1, characterized in that, Also includes: The aging degree of the system is assessed based on the cumulative blowing time and the number of adjustments. And when the aging degree exceeds the preset threshold, the safety optimization range in the improved target detection model is adjusted accordingly.

10. The online identification method for the end point of copper converter blowing based on an improved target detection model as described in claim 1, characterized in that, The improved target detection model is a neural network model trained based on historical blowing quality data and actual adjustment response data; The neural network model can adaptively optimize the combination of parameters of the feature extraction signal based on the deviation between the real-time monitored recognition state and the target state.

Images