Anode furnace oxidation process data management and end point true value calibration method

Abstract

This invention proposes a method for data governance and endpoint truth value calibration in the anode furnace oxidation process, belonging to the field of copper pyrometallurgical technology. The method includes: acquiring the operating state parameters and process signal data of the anode furnace; constructing a state gating signal based on the operating state parameters to identify effective and ineffective operating intervals; within the effective operating interval, calibrating the oxidation start and end times based on the physical response characteristics of the process signal data; identifying and reconstructing distorted signals in the process signal data; and calculating the endpoint truth label for each time step based on the state gating signal, oxidation start boundary, oxidation end boundary, and the reconstructed process signal data. This invention, through systematic data governance, constructs a high-quality training dataset with accurate start and end boundaries and physically interpretable labels from noisy and discontinuous raw DCS data, providing a reliable data foundation for intelligent prediction of the anode furnace oxidation endpoint.

Description

Technical Field

[0001] This invention belongs to the field of copper pyrometallurgical technology, and specifically relates to a method for data management and endpoint true value calibration of the anode furnace oxidation process. Background Technology

[0002] Anode refining is the final and crucial step in the pyrometallurgical copper smelting process, typically located at the end of a continuous copper smelting system consisting of three furnaces (top-blown furnace, side-blown furnace, and anode furnace). Its core process is divided into two stages: oxidation and reduction. During oxidation, compressed air (or oxygen-enriched air) is blown into the molten copper to oxidize impurities such as sulfur, lead, and arsenic, which are then removed in a gaseous state (e.g., sulfur dioxide) or through slag formation. During reduction, a reducing agent (e.g., natural gas, heavy oil, or ammonia) is introduced to remove excess oxygen from the molten copper, ultimately casting anode plates with acceptable chemical composition. The accuracy of determining the oxidation endpoint directly determines the quality of the anode plates (especially whether the sulfur content meets standards), the energy consumption level of the subsequent reduction stage, and the service life of the furnace lining refractory material.

[0003] In current production, the determination of the oxidation endpoint mainly relies on the operator's personal experience, such as observing the churning state of the molten pool (whether the copper surface shows a "blooming" phenomenon), taking out molten copper with a sampling spoon, and judging based on its surface morphology after cooling and solidification (such as whether there are "shrinkage cavities" or "protrusions"). This traditional method has three prominent problems: First, it is highly subjective. Operators with different shifts, years of service, and operating habits have inconsistent standards for judging the state of the same molten pool, resulting in significant fluctuations in product quality between shifts and furnaces, making it difficult to achieve production standardization.

[0004] Secondly, there are two risks. Insufficient oxidation will lead to excessive sulfur content, resulting in substandard physical properties of the anode plate (such as conductivity and ductility). When operators have difficulty accurately judging the oxidation endpoint, the blowing time may be appropriately extended to reduce the risk of insufficient oxidation. This may cause copper liquid to over-oxidize (the dissolved oxygen content in the copper liquid is too high), which may increase the consumption of reducing agent in the subsequent reduction stage (natural gas reacts with dissolved oxygen to produce CO2 and H2O), and may increase the chemical corrosion and thermal shock effects on the furnace lining refractory material, shorten the service life of the furnace lining, and increase the frequency of furnace shutdown for maintenance.

[0005] Third, it has a strong lag. Sampling and observation require a certain amount of operation time and reflect the melt state at the time of sampling. It is difficult to respond in real time to the rapidly changing reaction process in the furnace, especially in the accelerated stage of oxidation reaction, where the operator's judgment often lags behind the actual changes in the furnace state.

[0006] To overcome the drawbacks of manual judgment, the industry has begun exploring quantitative and reproducible automatic predictions of oxidation endpoints based on process data (i.e., real-time sensor data collected by the DCS distributed control system, such as flow rate, pressure, temperature, SO2 concentration, etc.) and machine learning models. The basic idea is to use historical furnace operation data as samples, and the remaining time until the end of oxidation at each time point as the prediction target (i.e., label), to train a temporal prediction model (such as a Long Short-Term Memory (LSTM) network, a Temporal Convolutional Network (TCN), or a Transformer architecture). This allows for real-time output of the predicted oxidation endpoint in actual production, providing operators with decision support or enabling closed-loop automatic control.

[0007] However, analysis of the actual operating conditions of the anode furnace and historical DCS data revealed that the validity of the aforementioned data-driven method does not depend on the selection of the model algorithm or hyperparameter tuning, but rather on the quality of the training data itself, especially the accuracy of the ground truth label for "oxidation start and end times." It is precisely in this crucial prerequisite for model training that existing technologies face numerous unresolved engineering challenges, specifically in the following four aspects: First, the start and end times of oxidation are difficult to accurately mark. Although there are manual oxidation feedback signals on site (operators mark the start and end of oxidation via buttons or human-machine interfaces), there may be significant deviations between these signals and the actual oxidation process. In some cases, the manual feedback signals may be delayed or recorded retroactively. If such manual signals are used directly as labels, they will introduce large time labeling errors, which may have a significant impact on model training based on second-level sampling data. At the same time, in a three-furnace system, crude copper from the upstream furnace continuously flows into the anode furnace through chutes. During the charging process, the total amount of molten copper and the sulfur content are constantly changing. If the moment of the first ventilation is simply taken as the starting point of oxidation, the mathematical premise of the endpoint prediction—that is, material conservation under the condition of constant total amount—is difficult to strictly hold because the total mass of molten copper in the furnace has not yet been determined (more crude copper will be added later). This will cause the remaining time labels established based on the assumption of constant total amount or material conservation to be biased, which may affect the training effect of the prediction model.

[0008] Second, the shut-off signal suffers from tailing distortion. The physical marker of the end of oxidation is shutting off the air supply or rotating the furnace (moving the furnace body out of the oxidation position). However, the compressed air flow signal collected by the DCS does not immediately return to zero after the valve is closed. Instead, there is a residual gas release process of about one to two minutes. After the valve is closed, the compressed air remaining in the air supply pipeline continues to be discharged into the melt, and its flow curve shows a sharp drop from a high level followed by a slow tailing to zero. If the moment when the flow reaches zero in the DCS data is directly taken as the oxidation endpoint, the invalid tailing period after the air supply is shut off will be included in the endpoint label, thus making the endpoint label include a time period that does not correspond to the actual effective oxidation process, and may reduce the effectiveness of the model training samples. In addition, there are differences in shutdown methods depending on the operating conditions: under high air volume conditions, emergency valve closure will cause a cliff-like drop (the flow rate drops from the peak to below 20% within a few seconds), while under low air volume conditions, gradual valve closure may show a steady decline in flow until it reaches zero. It is difficult to uniformly and adaptively identify these two significantly different shutdown modes by simply relying on a fixed threshold or a fixed slope.

[0009] Third, key chemical signals are contaminated by periodic noise. SO2 analyzers (usually based on non-dispersive infrared absorption or ultraviolet fluorescence) require periodic backflushing during operation to prevent smoke and dust from clogging the gas path. During backflushing (usually lasting tens of seconds to several minutes), the SO2 concentration signal exhibits pseudo-abnormal waveforms such as a numerical plateau (maintaining a fixed value), a sudden drop (dilution effect), and a rapid rebound. If these noises are not identified and distinguished and directly entered into the training data, they may interfere with the model's learning of SO2 concentration change patterns. Furthermore, real-world changes in operating conditions (such as sudden changes in sulfur content in the furnace caused by periodic slag removal and dumping operations, or changes in reaction rate caused by fluctuations in molten pool temperature) can also cause drastic changes in SO2 concentration. These two types of noise share certain similarities in numerical patterns, necessitating the establishment of an effective differentiation mechanism.

[0010] Fourth, key parameters such as temperature exist as discrete events rather than continuous signals. The temperature of the molten copper is not provided by a continuous online temperature measurement device (continuous temperature measurement technology for high-temperature melts faces engineering challenges such as short sensor lifespan and large drift), but is generated by intermittent insertion of a manual or automatic temperature gun into the melt. Therefore, the temperature data stored in the DCS consists of a mixture of discrete valid temperature values ​​(each measurement lasts several seconds, yielding a valid reading) and a large number of invalid low values ​​(room temperature or flue gas temperature readings generated when the thermocouple is exposed to air without the temperature gun being inserted into the melt). This data format is not suitable for direct input into the prediction model as continuous time-series features. It is necessary to accurately identify and extract the actual temperature measurement events from the invalid low-value data stream and reconstruct the timestamp and true temperature value of each temperature measurement event so that it can be used as process features for subsequent models.

[0011] Therefore, the anode furnace oxidation process exhibits complex characteristics, including intermittent and discontinuous operation (alternating stages of charging, oxidation, slag removal, and reduction), signal distortion (air stoppage and backflushing noise), and ambiguous start and end boundaries (oxidation start depends on the completion of charging, and end depends on cliff recognition rather than simple zeroing). Raw DCS data is difficult to directly use for training high-precision end-point prediction models. How to reliably determine the start and end times of oxidation, reconstruct distorted signals, and derive logically interpretable features from the aforementioned noisy, discontinuous, and ambiguous industrial data through systematic data governance methods, thereby constructing a high-quality training dataset and end-point ground truth labels, is a crucial prerequisite for the effective engineering application of intelligent prediction of anode furnace oxidation ends, and is also the core technical problem this invention aims to solve.

[0012] The following are two prior art patent documents in this field as examples: Prior art 1: "An intelligent system and method for judging the endpoint of copper oxidation-reduction in an anode furnace", publication number CN120635799A.

[0013] Brief Description: This technical solution proposes an endpoint determination method based on dual-modal image acquisition and a hybrid neural network. Its system architecture includes: a high-resolution microscope camera, dual-wavelength light sources (for providing illumination in different wavelengths), an air curtain dustproof device (for protecting the lens from smoke and dust contamination), and a polarizing filter (for eliminating reflective interference from the copper molten metal surface). During the oxidation phase, the copper sample image is segmented using dynamic thresholding in the HSV color space, and the porosity index in the image is calculated morphologically. During the reduction phase, the texture contrast of the sample surface is extracted using dual light sources. The model architecture employs a temporal convolutional network and a dual-path convolutional neural network (one path processes porosity temporal data, and the other path processes texture features), combined with a fuzzy PID controller and a digital twin model for feedback adjustment.

[0014] Objective analysis of existing technologies: This technical solution has the following limitations: ① High hardware dependence and poor on-site adaptability. This solution relies on precision optical hardware (microscope camera, dual-wavelength light source, air curtain device, polarizing filter). Under harsh working conditions such as high temperature (ambient temperature can reach above 60℃), high dust (high dust concentration, small particle size), and strong vibration (caused by furnace shaking operation and melt tumbling), the optical lens and precision adjustment mechanism are difficult to operate stably for a long time. Problems such as optical path deviation, lens contamination, and light source attenuation will lead to a decrease in image quality, thus affecting the reliability of endpoint determination. Moreover, the equipment maintenance cost is high. ② The data quality problem of DCS process signals is not solved. The endpoint determination of this solution is based on the "instantaneous observation" of the operator sampling and taking pictures of copper samples. It does not address how to calibrate the actual start and end times of oxidation from continuous DCS process signals (flow rate, pressure, SO2 concentration, etc.), nor does it solve the distortion problems unique to DCS data, such as the trailing of the air stop signal, backflush noise of the SO2 analyzer, and reconstruction of discrete temperature measurement signals. Therefore, this solution belongs to a different technical solution path from the present invention, and cannot cover the data governance needs targeted by the present invention.

[0015] Prior Art 2: "Intelligent Judgment Method for Oxidation-Reduction Endpoint of Anode Furnace Based on Computer Vision", Publication No. CN121010829A.

[0016] Brief Description: This technical solution focuses on dataset construction and cleaning methods. Images of both sides of copper samples are acquired using an industrial camera to establish a database containing furnace number and stage labels (oxidation stage, reduction stage). The model employs an "encoder + multilayer perceptron" (MLP) structure, outputting stage classification (determining whether the current stage is oxidation or reduction) and a predicted time to the end point. This solution particularly emphasizes the correction of manually predicted data—by having multiple operators independently make judgments and taking the statistical average, the random error of individual human judgments is mitigated.

[0017] An objective analysis of existing technologies reveals the following limitations: ① The label source is essentially a fit to human experience, failing to eliminate systematic biases in human judgment. The label construction in this scheme heavily relies on operator predictions of near-endpoint samples and label assignment based on simple linear time extrapolation (assuming a constant oxidation reaction rate). Essentially, it's using a machine learning model to fit human experience. If operator judgment exhibits systematic biases (e.g., a general tendency to prematurely determine the endpoint, or inaccurate judgments under specific operating conditions), the model will inevitably inherit and amplify these biases. ② Data cleaning targets image samples rather than continuous process signals. Its data cleaning methods focus on image quality dimensions such as image clarity, illumination consistency, and sample placement angle, without addressing issues like the calibration of start and end times for continuous process signals, removal of airflow tails, identification of SO2 concentration signal backflushing noise, and reconstruction of discrete temperature measurement events. Therefore, this scheme addresses a different problem domain than this invention—the former solves the problem of image sample labeling quality, while the latter addresses the problem of DCS continuous signal data quality.

[0018] In summary, in the prior art, Reference Document 1 relies on precision optical hardware for instantaneous image observation, while Reference Document 2 relies on manual experience-based labeling as the source of truth values ​​and involves cleaning of image samples. As can be seen from the above disclosures, the prior art mainly focuses on image acquisition, image feature extraction, and data annotation based on manual experience to determine the endpoint of anode furnace oxidation-reduction. However, no specific publicly available solutions have been found for data governance processes such as oxidation start-end boundary calibration, air stop tailing processing, SO2 anomaly identification, and discrete temperature measurement event reconstruction based on DCS continuous process signals. Therefore, compared to the prior art, the technical solution proposed in this invention has outstanding substantive features and significant technological advancements. Summary of the Invention

[0019] In view of the above situation, the main objective of this invention is to propose a method for data management and endpoint truth value calibration in the anode furnace oxidation process, so as to solve the above-mentioned technical problems.

[0020] This invention proposes a method for data management and endpoint truth value calibration in an anode furnace oxidation process, the method comprising the following steps: The operating status parameters and process signal data of the anode furnace are acquired and used as the status data and signal data to be processed, respectively. A state gating signal is constructed based on the operating condition parameters. This state gating signal is used to identify the effective operating range and ineffective operating range of the anode furnace. Within the effective operating range, based on the physical response characteristics of the process signal data, the start and end times of the anode furnace oxidation process are calibrated and used as the oxidation start boundary and oxidation end boundary, respectively. Perform distortion identification and reconstruction operations on distorted signals in process signal data to generate reconstructed process signal data; Based on the state gating signal, oxidation start boundary, oxidation end boundary, and reconstructed process signal data, calculate the endpoint truth label for each time step of the anode furnace oxidation process.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Eliminate manual feedback signals and reduce systematic bias in training labels from the source.

[0022] This invention uses the metallurgical process mechanism as the criterion (the logical intersection of charging state, oxidation air volume establishment, and furnace tilt angle) to calibrate the oxidation initiation boundary through multi-source signal fusion; and uses the physical response characteristics of compressed air flow (adjacent point ratio cliff identification) and furnace tilt angle as auxiliary arbitration to calibrate the oxidation termination boundary. The entire process does not rely on artificial oxidation feedback signals with large deviations and strong lags as the label source, thus reducing the impact of lag or deviation in artificial feedback signals on training labels from the data source, improving the consistency and repeatability of oxidation initiation and termination boundary calibration, and providing a more objective physical basis for the prediction targets of subsequent models.

[0023] 2. Adaptive cliff identification helps eliminate invalid trailing periods after wind cessation, improving the accuracy of endpoint boundary calibration.

[0024] This invention employs a reverse full-segment scanning method based on the ratio of adjacent points to identify airflow cliffs, and provides a unified and adaptive criterion for both high-level cliff shutdown and low-level stable zeroing modes—eliminating the need for manually setting fixed thresholds or slope limits to distinguish different operating conditions. Under high airflow conditions, this criterion can identify the cliff location based on the variation characteristics between adjacent airflow sampling points, thereby reducing the impact of the residual gas venting tail section after airflow shutdown on the oxidation termination boundary calibration. Under low airflow conditions, when no significant abrupt change is detected, the last non-zero sampling point is used as an auxiliary criterion, thereby improving the consistency of endpoint boundary calibration under different airflow shutdown modes, reducing the possibility of invalid time periods being included in the endpoint label, and substantially improving label accuracy.

[0025] 3. Distinguishing SO2 backflushing noise from actual operating conditions by duration helps reduce the interference of SO2 backflushing anomalies on process signals.

[0026] This invention addresses the pseudo-anomaly signals generated by the periodic backflushing of SO2 analyzers. It employs a sliding maximum value to determine a baseline, identifies anomaly intervals by calculating the deviation of the current signal value from the baseline, and uses "anomaly duration" as a classification criterion—short-term anomalies (not exceeding a preset duration threshold) are identified as sensor backflushing noise and interpolated for reconstruction, while long-term anomalies are identified as actual operating condition changes (such as slag dumping) and retained. Simultaneously, the anomaly intervals are cross-compared with a slag dumping state mask to further avoid misclassifying drastic SO2 changes caused by actual slag dumping operations as noise and incorrectly rejecting them. Compared to simple threshold filtering or fixed-pattern rejection, this method can jointly judge candidate anomaly intervals through anomaly duration and slag dumping state mask, helping to distinguish between sensor backflushing anomalies and actual operating condition changes. Interpolation reconstruction is performed on intervals identified as backflushing anomalies, thereby reducing the impact of anomaly signals on subsequent feature extraction and model input, while retaining signal information related to actual process changes, thus helping to reduce the interference of SO2 backflushing anomalies on process signals.

[0027] 4. Reconstruct discrete temperature measurement events to transform discontinuous signals into usable time-series features.

[0028] This invention addresses the engineering characteristic of copper molten temperature signals, which consist of a mixture of discrete valid temperature measurements and numerous invalid low values. It identifies temperature measurement events by eliminating invalid data below the lower limit of the melt temperature and detecting numerical step changes in the valid data. The occurrence time and subsequent value of each temperature change are recorded as a discrete temperature measurement event sequence. Through this processing, the invalid and mixed data, which could not be directly used as continuous temporal features, is restored into a temperature measurement event sequence with clear timestamps and physical meaning. This allows discrete temperature measurement information to participate in subsequent data processing and endpoint prediction modeling in the form of an event sequence, enabling temperature, a key process parameter, to serve as a valid input feature in subsequent endpoint prediction modeling.

[0029] 5. Construct end-point truth labels with process implications based on effective blowing time.

[0030] After completing signal purification and boundary calibration, this invention further derives physical characteristics such as oxygen standardized correction flow rate, kinetic stirring index, dynamic air-fuel ratio, and effective energy flow density. It uses the blowing state mask to accumulate the effective operating time, and maps the discrete operating periods of multiple furnace shaking interruptions to the effective blowing time axis. Based on this, it constructs the remaining effective blowing time label, so that the generated endpoint truth label has a clear process time meaning, which is convenient for subsequent endpoint prediction models.

[0031] 6. This invention is mainly based on the existing process signals and status parameters of the DCS system. It achieves data governance through rule-based processing such as status gating, boundary calibration, anomaly identification, signal reconstruction and label calculation. Without the need for additional dedicated acquisition hardware, it provides a processed training data foundation for the subsequent oxidation endpoint prediction model.

[0032] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by means of embodiments of the invention. Attached Figure Description

[0033] Figure 1 This is an overall flowchart of the data management and endpoint truth calibration method for the anode furnace oxidation process proposed in this invention; Figure 2 A schematic diagram of adaptive cliff identification and calibration for oxidation endpoint; Figure 3 Comparison of SO2 backflush noise identification and interpolation reconstruction before and after. Detailed Implementation

[0034] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0035] The following example uses actual production data from the anode furnace of a copper smelter, along with accompanying data. Figure 1-3 This invention provides a detailed description of the data processing and endpoint truth value calibration method for the anode furnace oxidation process. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0036] This embodiment discloses a method for data management and endpoint truth calibration in an anode furnace oxidation process, as shown in the attached figure. Figure 1 As shown, the method of the present invention includes six steps in sequence: multi-source data acquisition and state gating signal construction, adaptive calibration of oxidation start and end times, distortion signal reconstruction, feature derivation based on physical mechanisms, sample construction and label inverse calculation, and quality verification and consistency check. Each step will be explained in detail below with reference to specific implementation methods.

[0037] Step 1: Multi-source data acquisition and state gating signal construction; First, raw sensor data is read through the DCS interface to establish a mapping from equipment tag number to physical variables, thereby acquiring the operating status parameters and process signal data of the anode furnace. The operating status parameters include at least the anode furnace body tilt angle. Process signal data includes compressed air flow rate, pressure, and temperature for oxidation; natural gas flow rate and nitrogen flow rate for reduction; nitrogen flow rate in each branch of the permeable brick; oxygen and natural gas flow rate for dilute oxygen combustion; copper molten metal temperature; flue gas pressure and temperature; SO2 concentration; and feeding feedback signals, etc.

[0038] A state gating signal is constructed based on the furnace body tilt angle from the operating condition parameters. In this embodiment, based on process experience, an tilt angle of approximately 0° is considered a safe position, approximately 20° is the feeding position, approximately 45° is the oxidation-reduction position, and approximately 60° is the slag removal position. Accordingly, the state gating signal includes a blowing state mask. With slag dumping status mask The specific generation method is as follows: set the threshold value for the inclination angle of the blowing position. Compare the furnace body tilt angle with this threshold; if... If the condition is met, a blowing status mask value of 1 is generated, indicating that the current time is within the valid operating range; otherwise, a value of 0 is generated, indicating that the current time is within the invalid operating range. Similarly, a threshold value for the slag dumping position inclination angle is set. ,like If the condition is met, a slag dumping status mask value of 1 is generated, indicating that the current time is within the slag dumping operation range; otherwise, a value of 0 is generated, and the corresponding relationship is as follows: ; ; in, Let t be the inclination angle of the anode furnace body. The threshold for the inclination angle of the blowing position is set at 40°~50°. The threshold for the tilt angle of the slag dumping position. For the refining status mask, This is the mask for the slag removal status.

[0039] To eliminate interference from leakage readings in ineffective operating areas, a gating operation is performed on the gas flow characteristics in the process signal data. This forces the gas flow characteristics in the process signal data identified by the blowing status mask as ineffective operating areas to be set to zero, generating gating-processed gas flow data. Specifically, when the blowing status mask identifies an ineffective operating area (i.e., When the process signal data is processed, the gas flow characteristics (including compressed air flow, nitrogen flow, etc.) are forcibly set to zero to generate gated gas flow data. Simultaneously, the slag removal status marker is used in subsequent steps to distinguish the source of SO2 concentration signal anomalies—if a certain SO2 anomaly interval overlaps with the slag removal operation interval, it should be determined as a genuine change in operating conditions rather than a sensor spurious anomaly.

[0040] Step 2: Adaptive calibration of oxidation start and end times; This step does not use artificial oxidation feedback signals with large deviations. Instead, it uses physical response characteristics to calibrate the oxidation start boundary and oxidation end boundary within the effective working range constructed in step one.

[0041] (a) Determination of the oxidation initiation boundary; The premise for oxidation to enter a predictable convergence process is a fixed total amount of molten copper. Therefore, the start time is defined as the moment when feeding stops, the furnace enters the oxidation position, and the oxidation airflow is established. In specific operation, the oxidation airflow signal is extracted from the gas flow data after gating processing generated in step one, and the feeding feedback signal and furnace tilt angle signal are obtained simultaneously. The earliest of the following three conditions that are met simultaneously is determined as the start time: Condition 1: Feeding feedback signal is OFF (feeding stopped); Condition 2: The oxidation air volume is greater than the preset air volume threshold, which is 40% of the normal blowing air volume (it can be adjusted within the range of 30%-50%, and the middle value of 40% is used in this embodiment). Condition 3: The furnace body is tilted at the oxidation position, that is, the tilt angle is approximately 45° (an engineering tolerance of ±3° is allowed).

[0042] The earliest moment of the intersection of the above three conditions is recorded as the oxidation initiation boundary, and the corresponding relationship is as follows: ; in, This is the oxidation initiation boundary. The compressed air volume used for oxidation at time t. The air volume threshold used as the starting point for judgment is generally taken as 30% to 50% of the normal blowing air volume.

[0043] (ii) Determination of the oxidation termination boundary; The physical marker for the end of oxidation is the shutdown of the airflow or the shaking of the furnace. However, the oxidation airflow signal exhibits a residual gas release tail for approximately two minutes after the valve is closed, and there are two modes: a high-level cliff-like shutdown and a low-level direct zeroing. This embodiment employs an adaptive cliff-like shutdown method based on the ratio of adjacent points, combined with an inclination angle-assisted criterion, to achieve accurate calibration.

[0044] Main criterion (airflow cliff identification): Extract the time series data of the oxidation air volume signal after the airflow is stopped from the gated gas flow data generated in step one, and use it as the air volume sequence to be identified. Calculate the air volume ratio between adjacent sampling points in the air volume sequence to be identified sequentially along the reverse time direction (i.e., from the tail to the head), generating a reverse adjacent point ratio sequence. The corresponding relationship is as follows: ; in, and The air volume is the air volume of adjacent sampling points in the air volume sequence to be identified. This is the value at time t in the reverse adjacent point ratio sequence; Set the ratio threshold ρ to 0.70 (this can be adjusted within the range of 0.65 to 0.75; in this embodiment, it is set to 0.70). In the reverse adjacent point ratio sequence, filter out all points that meet the criteria... The trigger point. To accurately capture the leading edge of the cliff and eliminate residual gas release trails, the trigger point is located... The moment with the minimum value (i.e., the largest reverse decline, corresponding to the most drastic forward deceleration) is taken as the core moment of the mutation. The corresponding relationship is as follows: ; Using the core moment of this mutation as a benchmark, continue tracing backwards in time until the ratio between two adjacent points recovers and stabilizes within a stable range (i.e., And at least three sampling points are continuously used to filter out instantaneous signal jitter or noise interference, ensuring the stability of the airflow establishment state. This refers to the high-level turning point at the leading edge of the cliff step before entering this stable range. The corresponding time is determined to be the end time of the anode furnace oxidation process. This allows for the precise removal of the trailing section of residual gas after the cliff.

[0045] To more clearly illustrate the adaptive advantages of the above ratio criterion compared to fixed threshold or fixed slope criteria, the following numerical examples with two typical working conditions are provided: High-volume cliff shutdown condition: If the normal blowing air volume before shutdown is 660 Nm 3 / h, after the valve is closed, the air volume suddenly drops to 320 Nm 3 / h, then the air volume ratio of adjacent sampling points at the core of the cliff is about 320 / 660≈0.48. This value is significantly less than the preset ratio threshold ρ=0.70. By scanning the entire section in reverse, the point with the minimum ratio (i.e. the point of the most drastic deceleration) can be accurately located as the core moment of the sudden change. Then, the sampling point before this moment is taken as the high-level turning point at the leading edge of the cliff. The residual gas release tail section about two minutes after the cliff is accurately cut off, and the oxidation end boundary is located to the actual physical moment when the effective oxidation reaction terminates.

[0046] Low airflow stable zeroing condition: If the airflow value before shutdown was 250 Nm 3 / h, after the valve is closed, the airflow decreases smoothly in a pattern of 250→250→0 until it reaches zero (i.e., the ratio of adjacent points is 1 or infinity, without significant abrupt changes). Therefore, there is no sequence of reverse adjacent point ratios that satisfies this condition. The trigger point is not triggered by the cliff recognition logic. Instead, the moment corresponding to the last non-zero point in the non-zero sequence (i.e., the last valid reading before returning to zero) is directly taken as the oxidation end boundary.

[0047] Therefore, the same set of criteria based on the ratio of adjacent points can adaptively handle the two completely different shutdown conditions without manually switching thresholds or setting different modes. This avoids the problems of the fixed threshold method mistakenly including the trailing section in the effective range under high air volume conditions or missing the endpoint due to the low signal amplitude under low air volume conditions.

[0048] Auxiliary Criterion 1 (Smooth Zeroing Mode): Under low airflow conditions, if the airflow sequence decreases steadily until it reaches zero (i.e., the cliff condition of Wt<ρ is not triggered in the entire reverse sequence), then the time corresponding to the last non-zero sampling point in the airflow sequence to be identified is determined as the oxidation end boundary.

[0049] Second auxiliary criterion (inclination angle arbitration): Obtain the furnace body tilt angle signal of the anode furnace. When the oxidation air volume signal does not show the above-mentioned abrupt change, but the furnace body tilt angle undergoes a step change and turns out of the oxidation position (i.e. the tilt angle drops rapidly from about 45° to below 20°), in order to correct the endpoint determination in the scenario where the air volume signal does not change abruptly, the time corresponding to the step drop edge of the furnace body tilt angle is determined as the oxidation end boundary.

[0050] It should also be noted that the delayed venting of residual gas in the pipeline causes the SO2 concentration signal to return to zero several minutes later than the actual endpoint. Therefore, the SO2 zeroing time is not used as the basis for endpoint determination, but only as a cross-validation reference. (Appendix) Figure 2 Taking the compressed air flow curve as an example, the core principle of the above adaptive cliff identification is illustrated: In the air volume sequence to be identified after the air stops, the cliff cut-off point (i.e. the core moment of the sudden change) is locked by scanning the ratio of adjacent points in reverse. The section after the cliff cut-off point is marked as "residual gas release tailings" and cut off. The moment before the leading edge of the cliff is taken as the oxidation end boundary, thereby accurately eliminating invalid tailing periods.

[0051] Through the above calibration, the accurate oxidation initiation boundary for each furnace batch was obtained. and oxidation end boundary This serves as the benchmark for subsequent truth value inversion.

[0052] Step 3: Distorted signal identification and reconstruction; (I) Identification and reconstruction of SO2 concentration backflush noise: SO2 analyzers require periodic backflushing for cleaning during operation. During backflushing, the signal may exhibit pseudo-anomalies such as a numerical plateau, a sudden drop, and a rapid recovery. This embodiment performs the following distortion identification and reconstruction operations on the sulfur dioxide concentration signal: First, the baseline is determined using the sliding maximum value. The time window width is set to 7 minutes (adjustable between 6-8 minutes; 7 minutes is used in this embodiment to ensure the local baseline accurately reflects the concentration level before backflushing, avoiding the baseline being dragged down by low values ​​during the backflushing process due to an excessively short window). For each time point t, the maximum sulfur dioxide concentration within that window is calculated, generating a local baseline sequence. .

[0053] Next, calculate the measured value of sulfur dioxide concentration at the current moment. The difference between the value and the corresponding value in the local baseline sequence is used to generate a difference sequence, and the corresponding relationship is as follows: ; in, The value at time t in the difference sequence. The maximum sliding value at time t. Let be the measured SO2 concentration at time t. Set a difference threshold. (This embodiment uses 20% of the normal SO2 concentration range, which can be calibrated according to actual operating conditions.) When When the time interval is reached, the corresponding continuous time period is marked as a candidate anomaly interval. The duration of each candidate anomaly interval is then obtained. .

[0054] To distinguish between sensor false anomalies and actual changes in operating conditions, the duration of candidate anomaly intervals is determined. Compared with the preset duration threshold In comparison, considering that the actual backflush cycle of the SO2 analyzer in the field is usually 5-6 minutes, Set the time to 7 minutes (adjustable between 6-8 minutes) to ensure a complete backflushing process is covered, avoiding misinterpreting actual backflushing noise as a change in operating conditions. When the candidate abnormal interval is identified as the backflush noise interval, linear interpolation reconstruction is performed on the sulfur dioxide concentration signal within this backflush noise interval to generate the reconstructed sulfur dioxide concentration signal; when When the actual operating conditions change (such as drastic changes in SO2 concentration caused by process operations such as slag dumping and slag discharge), the original sulfur dioxide concentration signal within that range is retained.

[0055] Furthermore, when determining whether a candidate abnormal interval belongs to the backflushing noise interval, the candidate abnormal interval is compared with the slag dumping state mask generated in step one. When the candidate abnormal interval overlaps with the slag dumping operation interval, to avoid misjudging the actual operating condition change as noise, the candidate abnormal interval is directly determined as the actual operating condition change interval, and the original sulfur dioxide concentration signal within that interval is retained. (Appendix) Figure 3The upper part shows the backflush noise segment present in the original SO2 concentration signal (manifested as a pseudo-abnormal waveform with a sudden drop, plateau, and rebound in signal). The lower part shows the signal morphology after the above identification and linear interpolation reconstruction—the backflush noise segment is replaced by a smooth and continuous interpolation curve, which naturally connects with the normal signals before and after, restoring the true trend of SO2 concentration change.

[0056] (II) Reconstruction of discrete signals of copper molten temperature: The temperature of molten copper is composed of discrete effective values ​​from intermittent temperature measurements mixed with a large number of invalid low values, making it difficult to use directly as a continuous feature. This embodiment performs the following reconstruction operation on the molten copper temperature signal: First, invalid data in the copper molten temperature signal that is lower than a preset temperature threshold (1000℃ in this embodiment) is removed to generate a purified copper molten temperature sequence. This threshold is set based on the melting point of copper molten metal and the minimum operating temperature of the process. Readings below this value are usually spurious values ​​generated when the thermocouple is not inserted into the melt or when there is poor contact.

[0057] Next, to restore discrete invalid low values ​​to real continuous temperature measurement events, step changes in the temperature sequence of the purified copper solution are detected. Specifically, the detection method is as follows: scanning forward in time, once the absolute difference between the current temperature value and the previous temperature value exceeds a preset step detection threshold (5℃ in this embodiment, which is greater than the range of conventional measurement noise fluctuations), it is considered a new temperature measurement event. Whenever a step change is detected, the start time of the change and the subsequent temperature value are recorded, generating a discrete temperature measurement event sequence. Each event in this discrete event sequence corresponds to a real manual or automatic temperature measurement action, with its timestamp representing the actual time of the temperature measurement and its value representing the actual reading of that measurement.

[0058] Step 4: Feature derivation based on physical mechanisms; Within the pure oxidation window after calibration and reconstruction in steps two and three, the following time-series characteristics are derived based on the principles of metallurgical thermodynamics and kinetics, providing physically interpretable intermediate variables for the cumulative calculation of the truth label in step five: (1) Oxygen standardization correction: The compressed air volumetric flow rate is restored to the standard state volumetric flow rate according to temperature and pressure. The corresponding relationship is as follows: ; in, The standard state volumetric flow rate, To measure the compressed air flow rate under actual operating conditions, The actual measured compressed air gauge pressure (unit: kPa). The measured temperature of compressed air (unit: °C).

[0059] (2) Kinetic stirring index. Calculate the total melt stirring flux, which characterizes the intensity of gas-liquid turbulence in the molten pool. The corresponding relationship is as follows: ; in, This represents the total melt stirring flux. For the flow rate of compressed air used in oxidation, This represents the total flow rate of nitrogen.

[0060] (3) The dynamic air-fuel ratio characterizes the real-time stoichiometric conditions of the oxidation reaction, and the corresponding relationship is as follows: ; in, For dynamic air-fuel ratio, Let t be the standardized compressed air flow rate. Let be the natural gas flux at time t. ε It is a very small positive number that prevents the denominator from being zero.

[0061] (4) Effective energy flow density, which characterizes the impact work intensity of compressed air injected into the melt through the vent. Since the cross-sectional area of ​​the vent is constant, it can be simplified to the product of pressure and standard flow rate. The corresponding relationship is as follows: ; in, For effective energy flux density, For compressed air pressure, To standardize compressed air flow rate.

[0062] (5) Cumulative net oxygen consumption: The compressed air flow component in the gated gas flow data generated in step one, after being corrected for oxygen standardization as described above, is used as the accumulation item in the point-by-point accumulation calculation of cumulative net oxygen consumption. Taking the oxidation initiation boundary as the starting point for accumulation, the effective time period for accumulation is selected using the blowing state mask, and the reaction progress is characterized by point-by-point accumulation. Its discrete accumulation expression is as follows, and the corresponding relationship is as follows: ; in, Oxidation starting point By time t ( The cumulative net oxygen consumption, Let i be the standardized compressed air flow rate at the i-th sampling point. This is the blowing state mask for the i-th sampling point. The sampling interval is denoted as .

[0063] (6) Cumulative net blowing time: based on the oxidation initiation boundary Starting from the effective operating time, only the time spent working is accumulated to eliminate the timing discontinuity caused by furnace shaking interruptions. The corresponding relationship is as follows: ; in, Oxidation starting point The cumulative effective blowing time up to time t, excluding the period of interruption during furnace shaking.

[0064] Of the above characteristics, the cumulative net blowing time It is the most critical intermediate product, which connects the discrete time periods within the effective working range marked in step one into a continuous reaction progress, providing a cumulative benchmark with clear physical meaning for the back calculation of the truth label in step five.

[0065] Step 5: Sample construction and inverse calculation of endpoint truth labels; This step calculates the endpoint truth label for each time step based on the processing results of steps one through four above.

[0066] First, based on the state gating signal (specifically, the blowing state mask) generated in step one, the effective operating time between the oxidation start boundary and the oxidation end boundary is accumulated to generate the total effective blowing time for this heat. This refers to the cumulative time actually used for blowing during the oxidation stage of this furnace (excluding the period of interruption during furnace shaking).

[0067] Secondly, for any target time step t between the oxidation start boundary and the oxidation end boundary, the effective operating time between the oxidation start boundary and the target time step is accumulated, and the accumulated effective blowing time of the target time step is generated (the accumulation is also controlled by the blowing state mask, and the period of furnace shaking interruption is excluded).

[0068] Finally, the difference between the total effective refining time and the accumulated effective refining time is calculated, and this difference is used as the end-point truth label for the target time step: ; in, The truth label for the endpoint of the current target time step. This is the cumulative net blowing time at the oxidation endpoint, i.e., the total effective blowing time of the furnace.

[0069] This difference represents the effective net blowing time (in seconds) required from the current moment until the end of oxidation, and is the direct training target for the subsequent oxidation endpoint prediction model. During furnace shaking interruption ( During this period, due to the lack of progress in the timeline, It remains unchanged.

[0070] Furthermore, to focus on the critical stages of the oxidation reaction in its later stages and improve the model's prediction accuracy, data from non-critical stages—such as the initial reaction phase, when the melt is not yet fully homogenized, or when the reaction rate has not yet reached a steady state—can be removed. For example, data points with remaining time greater than 50%-80% (preferably 70%) of the total effective blowing time can be removed. This stage typically involves charging and heating processes, and the actual focus on endpoint prediction is relatively low. It should be noted that this screening step aims to eliminate interference from early non-steady-state data on model training and is not a necessary limitation of the present invention. In practice, this percentage threshold can be adjusted or even eliminated based on specific furnace conditions, raw material conditions, or model convergence. The remaining samples after the above optional screening are divided by furnace number using a leave-out method. Approximately 20% of the complete furnaces are used as an independent test set, and the remaining 80% are used as a training set, ensuring that there is no information leakage between the training set and the test set for time steps within the same furnace.

[0071] Step Six: Dataset Quality Verification and Consistency Check; After completing all the above processing, perform the following quality checks and consistency checks on the constructed dataset to ensure the reliability of the output data: (1) Time continuity check: Check whether there are unexpected jumps or breaks in the time series of each furnace, and ensure that the time interval between adjacent sampling points is a fixed sampling period Δt. If there is an abnormal interval, mark it and resample.

[0072] (2) Logical consistency check: Check the endpoint truth label Does it satisfy the monotonically non-increasing property (i.e., within the effective operating range, the remaining net blowing time should gradually decrease as time progresses, without any reverse increase)? If it does... If the growth is reversed, then backtrack to check the accuracy of the blowing state mask or whether the calibration boundary is reasonable, and if necessary, re-execute the boundary calibration in step two.

[0073] (3) Boundary consistency check: Check the oxidation initiation boundary. place, It should be equal to the total effective blowing time. At the oxidation termination boundary place, It should equal 0. If it does not, the calibration is considered a failure, and the data for that furnace should be excluded or recalibrated.

[0074] Through steps one through six above, a high-quality training dataset is obtained from the noisy, discontinuous raw DCS data. This dataset contains accurate start and end boundaries, purified signals and physically derived features, and labels calculated using a mechanism. This dataset can be directly used to train oxidation endpoint prediction models, eliminating reliance on manual experience or easily disturbed instantaneous image observations.

[0075] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for data management and endpoint truth calibration in an anode furnace oxidation process, characterized in that, The method includes the following steps: The operating status parameters and process signal data of the anode furnace are acquired and used as the status data and signal data to be processed, respectively. A state gating signal is constructed based on the operating condition parameters, and the effective and ineffective operating areas of the anode furnace are identified based on the state gating signal. Within the effective operating range, based on the physical response characteristics of the process signal data, the start and end times of the anode furnace oxidation process are calibrated and used as the oxidation start boundary and oxidation end boundary, respectively. Perform distortion identification and reconstruction operations on distorted signals in process signal data to generate reconstructed process signal data; Based on the state gating signal, oxidation start boundary, oxidation end boundary, and reconstructed process signal data, calculate the endpoint truth label for each time step of the anode furnace oxidation process.

2. The method for data management and endpoint truth calibration of the anode furnace oxidation process according to claim 1, characterized in that, Operating condition parameters include the furnace body tilt angle of the anode furnace; The state gating signals include the blowing state mask and the slag dumping state mask; Based on the comparison between the furnace body tilt angle and the preset blowing position tilt angle threshold, a blowing state mask is generated. The blowing state mask is used to identify the effective operating area. Based on the comparison between the furnace body tilt angle and the preset slag dumping position tilt angle threshold, a slag dumping status mask is generated. The slag dumping status mask is used to identify the slag dumping operation range. To eliminate interference from leakage readings in ineffective operating areas, a gating operation is performed on the gas flow characteristics in the process signal data. This forces the gas flow characteristics in the process signal data that are identified as ineffective operating areas by the blowing status mask to be set to zero, thereby generating gating-processed gas flow data.

3. The method for data management and endpoint truth calibration of the anode furnace oxidation process according to claim 1, characterized in that, Methods for defining the oxidation initiation boundary include: Extract the oxidation air volume signal from the gas flow data after gating; Acquire feeding feedback signals and furnace tilt angle signals; The moment when feeding stops, the oxidation air volume is greater than the preset air volume threshold, and the furnace body tilt angle is in the oxidation position is determined as the start time of the anode furnace oxidation process.

4. The method for data management and endpoint truth calibration of the anode furnace oxidation process according to claim 1, characterized in that, Methods for defining the oxidation termination boundary include: Extract the time series data of the oxidation air volume signal after the air stop from the gas flow data after gating, and use it as the air volume sequence to be identified; Calculate the air volume ratio between adjacent sampling points in the air volume sequence to be identified sequentially along the reverse time direction to generate a reverse adjacent point ratio sequence; To accurately capture the leading edge of the cliff and eliminate the trailing gas from the residual gas release, the core moment of the mutation that meets the preset mutation conditions is located in the reverse adjacent point ratio sequence, and the previous sampling moment corresponding to the core moment of the mutation is determined as the end moment of the anode furnace oxidation process.

5. The method for data management and endpoint truth calibration of the anode furnace oxidation process according to claim 4, characterized in that, Other methods for defining the oxidation termination boundary include: When there is no ratio value in the reverse adjacent point ratio sequence that meets the preset mutation condition, the time corresponding to the last non-zero sampling point in the air volume sequence to be identified is determined as the end time of the anode furnace oxidation process.

6. The method for data management and endpoint truth calibration of the anode furnace oxidation process according to claim 4, characterized in that, Other methods for defining the oxidation termination boundary include: Obtain the furnace body tilt angle signal of the anode furnace; When the oxidation air volume signal does not change abruptly, but the furnace body tilt angle changes abruptly and turns out of the oxidation position, in order to correct the endpoint determination in the scenario where the air volume signal does not change abruptly, the time corresponding to the falling edge of the furnace body tilt angle step is determined as the end time of the anode furnace oxidation process.

7. The method for data management and endpoint truth calibration of the anode furnace oxidation process according to claim 1, characterized in that, Process signal data includes sulfur dioxide concentration signals; Perform distortion identification and reconstruction operations on the distorted signal, including: Based on the maximum sliding value of the sulfur dioxide concentration signal within a preset time window, a baseline is determined, and a local baseline sequence is generated. Calculate the difference between the current sulfur dioxide concentration value and the corresponding value in the local baseline sequence, and generate a difference sequence; When the current difference in the difference sequence exceeds the preset difference threshold, the corresponding continuous time period is marked as a candidate abnormal interval; Obtain the duration of the candidate anomaly interval; To distinguish between sensor pseudo-anomalies and actual changes in operating conditions, when the duration is less than or equal to a preset duration threshold, the candidate anomaly interval is determined as the backflush noise interval, and the sulfur dioxide concentration signal within the backflush noise interval is reconstructed by linear interpolation to generate the reconstructed sulfur dioxide concentration signal. When the duration exceeds the preset duration threshold, it is determined to be a real operating condition change interval, and the original sulfur dioxide concentration signal within the real operating condition change interval is retained.

8. The method for data management and endpoint truth calibration of the anode furnace oxidation process according to claim 7, characterized in that, Methods for determining whether a candidate anomaly interval belongs to the backflush noise interval also include: Compare the candidate abnormal intervals with the slag dumping status mask; When the candidate abnormal interval overlaps with the slag dumping operation interval, in order to avoid misjudging the actual operating condition change as noise, the candidate abnormal interval is directly determined as the actual operating condition change interval, and the original sulfur dioxide concentration signal in the actual operating condition change interval is retained.

9. The method for data management and endpoint truth calibration of the anode furnace oxidation process according to claim 1, characterized in that, Process signal data includes copper molten metal temperature signals; Performing distortion identification and reconstruction operations on distorted signals also includes: Invalid data below a preset temperature threshold in the copper molten temperature signal are removed to generate a purified copper molten temperature sequence. To restore discrete invalid low values ​​to real continuous temperature measurement events, numerical step changes are detected in the temperature sequence of purified copper water. Whenever a numerical step change is detected, the start time of the change and the value after the change are recorded to generate a discrete temperature measurement event sequence.

10. The method for data management and endpoint truth calibration of the anode furnace oxidation process according to claim 1, characterized in that, Methods for calculating the endpoint truth label at each time step include: Based on the status gating signal, the effective working time between the oxidation start boundary and the oxidation end boundary is accumulated, and the total effective blowing time of the current furnace is generated. For any target time step between the oxidation start boundary and the oxidation end boundary, accumulate the effective working time from the oxidation start boundary to the current target time step, and generate the accumulated effective blowing time for the current target time step; Calculate the difference between the total effective blowing time and the accumulated effective blowing time, and use the difference in effective blowing time as the endpoint truth label of the current target time step.

Images