Smelting process data quality improvement and reconstruction method based on process event modeling

Abstract

The present application provides a smelting process data quality improvement and reconstruction method based on process event modeling, belonging to the field of smelting intelligence. The method comprises: obtaining sensor data and process event information; identifying the current process event and generating a process event label; matching a target interference model from a set of interference models based on the label to strip structured interference and obtain a de-interference signal; under the constraint of smelting mechanism equation, reconstructing the de-interference signal to extract core state signals conforming to physical laws; based on a preset process anchor point, aligning the time sequence and unifying the features of different data sources to generate a consistent process data set. The present application improves data processing to the process semantic level, solving the problems of structured interference stripping, signal physical distortion and data inconsistency.

Description

Technical Field

[0001] This invention belongs to the field of intelligent smelting, and specifically relates to a method for improving and reconstructing smelting process data quality based on process event modeling. Background Technology

[0002] In the smelting processes of metals such as copper, lead, and zinc, procedures like side-blowing smelting, multi-lance blowing, and anode refining generate a large amount of time-series data, such as molten pool temperature, furnace pressure, lance displacement, and flue gas composition. This data forms the core foundation for building artificial intelligence (AI) models for endpoint prediction and energy consumption optimization. However, the data quality at the smelting site is severely hampered by complex process operations, manifested in several ways: frequent process events (such as lance raising and lowering, backflushing, and slag discharge) introduce structured interference that overlaps with core signals; and data systems for different processes operate independently, leading to historical data time misalignment and feature loss. These problems severely restrict the training effect and prediction accuracy of downstream AI models. Therefore, how to systematically improve process data quality in industrial scenarios with strong process coupling and coexisting non-stationary interference is a key technical challenge that urgently needs to be addressed in this field.

[0003] Currently, several patented technologies have been proposed for industrial data governance, with representative solutions as follows: One approach is the "Industrial Time Series Data Governance Method" disclosed in CN120873555A. This method uses a large language model to parse the user's input governance requirements, combines them with a physical mechanism model, and plans a data governance path constrained by physical laws through an improved AI algorithm. Then, it calls governance tools such as filters to perform the processing. This approach achieves a preliminary combination of user intent and physical mechanism, but it is essentially still a "passive response" governance approach. The governance actions rely on external commands to trigger, and it cannot actively and in real time perceive high-frequency process events occurring on-site. Moreover, the parameters of the governance tools are fixed once set, making it difficult to dynamically model and accurately remove non-stationary disturbances (such as exponentially decaying pulses caused by backflushing) whose morphology is strongly correlated with process parameters.

[0004] Another approach is the "Intelligent Noise Detection Method for Industrial Sensors" disclosed in CN120974290A. This method constructs a sensor knowledge graph, embeds it into a pre-trained model via TransE, and fine-tunes the model using multi-source data, ultimately outputting the location, type, and confidence level of the noise. While this approach introduces domain knowledge to assist noise detection, its knowledge graph stores static rules, making it difficult to encode interference patterns with specific dynamic waveforms and time-varying parameters, such as "spray gun lifting." More importantly, its functionality is limited to the "detection" level, only outputting noise labels and failing to further generate quantitative waveforms of the interference signal and extract and reconstruct them from the original signal. Subsequent steps still rely on traditional filtering, inevitably leading to information loss in the core signal.

[0005] In summary, the existing technical solutions have three common defects, which constitute the core technical problem that this invention aims to solve: First, structured interference under frequency domain aliasing cannot be accurately separated. Whether it's AI-based governance path planning or knowledge graph-based noise detection, the denoising methods ultimately rely on general signal processing methods such as wavelet transform and low-pass filtering. These methods only consider the frequency characteristics of the signal. When structured interference caused by process operations (such as lance displacement pulses) overlaps with core smelting signals (such as molten pool temperature trends) in the frequency domain, they cannot be effectively separated, resulting in incomplete interference removal or damage to useful signals.

[0006] Second, the extraction of core signals lacks effective mechanistic constraints. Existing data-driven methods, after filtering out interference, rely solely on statistical optimality criteria to estimate core state signals (such as temperature), failing to incorporate fundamental thermodynamics, material balance, and other smelting mechanisms as hard constraints. Under highly non-stationary conditions, purely data-driven estimations are prone to physical distortions, such as temperature jumps that violate the law of heat conservation, which can seriously mislead downstream process decisions. How to mathematically and practically integrate mechanistic equations with signal estimation algorithms remains a challenging engineering problem.

[0007] Third, there is a lack of automatic consistency mechanisms for multi-source historical data. The entire smelting process involves multiple steps such as batching, smelting, and refining. Each step's DCS, MES, and testing systems record data independently, leading to misaligned timestamps and inconsistent feature dimensions for the same furnace in different systems. Existing solutions focus on processing individual data curves and lack batch consistency methods based on process logic that can automatically align the timeline and complete missing features, making it difficult to directly use valuable historical data for model training.

[0008] The root cause of the aforementioned shortcomings lies in the fact that existing technologies treat smelting data processing as a "pure signal" or "pure data" problem, rather than elevating it to the level of "process semantics." Therefore, the key to overcoming current technological bottlenecks lies in how to establish an explicit mapping between process events and signal interference, how to effectively integrate smelting mechanisms into the signal reconstruction process, and how to utilize process anchor points to achieve automatic alignment of multi-source data. Summary of the Invention

[0009] In view of the above situation, the main objective of this invention is to propose a method for improving and reconstructing the quality of smelting process data based on process event modeling, so as to solve the above-mentioned technical problems.

[0010] This invention proposes a method for improving and reconstructing smelting process data quality based on process event modeling. The method includes the following steps: Acquire sensor data during the smelting process, process event information reflecting process operation, and process mechanism parameters; Based on sensor data and process event information, identify the currently occurring process events and generate corresponding process event tags; Based on process event labels, target interference models matching the process events are determined from a pre-set set of interference models; By using a target interference model, structured interference is removed from sensor data to obtain a de-interference signal; Based on process mechanism parameters and under the constraints of the smelting mechanism equation, the interference removal signal is reconstructed to extract the core state signal that conforms to physical laws, so as to realize data reconstruction. Based on preset process anchor points, historical data from different data sources are time-series aligned and feature dimensions are unified to generate a consistent process dataset, thereby improving data quality.

[0011] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Achieved precise removal of structured interference under frequency domain aliasing; This invention establishes an explicit mapping from process events to interference waveform parameters by constructing a "process event semantic perception layer" and an "interference model set." Unlike general filtering methods that rely on signal frequency domain features, this invention identifies process event labels such as "backflushing" and "spray gun lifting" in real time, calls the corresponding parameterized interference models (such as exponentially decaying oscillation functions and damped step response functions), directly synthesizes interference components in the time domain, and performs subtraction operations. This method can accurately remove structured interference that overlaps with the core signal frequency domain, avoids accidental damage to useful signals in the frequency domain, and significantly improves signal fidelity.

[0012] 2. Ensured the physical robustness of core state signal extraction; This invention employs a two-stage architecture of "standard Kalman filtering + mechanistic consistency projection," using the smelting mechanism equation characterizing the heat accumulation relationship as a post-processing projection constraint. Under the premise of rigorous mathematical feasibility, this design ensures that the extracted core state signals, such as molten pool temperature, are both faithful to real-time measurements and strictly conform to fundamental thermodynamic laws, avoiding physical distortions that may arise from purely data-driven methods. Simultaneously, by using an adaptive regulator to re-estimate mechanistic parameters such as thermal efficiency online, it solves the problem of mechanistic model mismatch caused by changes in operating conditions during long-term operation, enhancing the method's adaptability to complex operating conditions.

[0013] 3. It has achieved automated, batch, and consistent processing of multi-source historical data; This invention defines "process anchor points" and constructs a cross-system semantic mapping table to automatically find the temporal correspondence between different data sources through event sequence context matching, and achieves point-by-point time axis correction based on a local offset function. Combined with missing feature imputation based on process stages and a segmented resampling strategy based on events, this invention automates and batches the manual furnace-by-furnace verification process, significantly improving the feature completeness and temporal consistency of historical data, and providing a high-quality training sample set for downstream AI models.

[0014] 4. Provide a data foundation for improving the performance of downstream AI models; The process dataset processed by the method of this invention possesses the characteristics of "low residual structured interference, high physical consistency, and cross-process dimension uniformity." Compared with directly using the original data containing interference, physical contradictions, and temporal misalignments, this high-quality dataset can reduce model overfitting and prediction bias caused by data noise and feature inconsistencies from the source, providing a solid data foundation for improving the generalization ability and prediction accuracy of downstream AI models such as endpoint prediction and energy consumption optimization.

[0015] 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

[0016] Figure 1 This is a flowchart of the smelting process data quality improvement and reconstruction method based on process event modeling proposed in this invention; Figure 2 This is the overall architecture diagram of the smelting process data quality improvement and reconstruction method based on process event modeling proposed in this invention. Detailed Implementation

[0017] 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.

[0018] These and other aspects of the embodiments of the present invention will become clear from the following description and accompanying drawings. In these descriptions and drawings, some specific embodiments of the present invention are specifically disclosed to illustrate some ways of implementing the principles of the embodiments of the present invention; however, it should be understood that the scope of the embodiments of the present invention is not limited thereto.

[0019] Example 1 Please see Figure 1This embodiment provides a method for improving and reconstructing smelting process data quality based on process event modeling. It aims to achieve a "process semantic-driven" data processing paradigm, solving the challenges of removing structured interference, extracting core state signals with physical fidelity, and ensuring consistency of historical data in smelting process data. Specifically, the method is as follows: Step S1: Acquire multi-source data during the smelting process; Acquire sensor data during the smelting process, process event information reflecting process operation, and process mechanism parameters.

[0020] Sensor data includes, but is not limited to, time-series data such as temperature (e.g., molten pool temperature), pressure (e.g., furnace pressure, duct pressure), displacement (e.g., spray gun height), and composition (e.g., SO2 concentration in flue gas). This data comes from field control systems such as DCS and PLC, and is typically sampled at frequencies of 1Hz or even higher.

[0021] Process event information refers to recorded data that can directly or indirectly characterize the occurrence of a specific process operation. This includes two types of sources: first, digital operation event logs in the PLC / DCS system, such as the change record of the "spray gun lifting start" switch signal, which contains the precise event name and trigger timestamp; second, operational indications that can be identified from sudden changes in sensor analog signals, such as when the displacement sensor value changes continuously and rapidly, it can serve as an auxiliary basis for judging the spray gun action.

[0022] At the same time, process mechanism parameters, such as time-varying oxygen flow rate, raw material ratio, and furnace ID, are also obtained for subsequent mechanism constraints and process correlation.

[0023] Step S2: Identify process events and generate process event tags; This step enables automatic perception from raw data to process semantics. The process is divided into two stages: "offline construction" and "online application".

[0024] Offline Construction Phase: Before the system goes live, process experts collect a large amount of historical furnace data covering various typical operating conditions. They manually review sensor trend curves and event logs, marking the precise times of various process events (such as lance lifting, backflushing, and slag discharge) and their corresponding signal waveform segments. For the waveform segments of each type of event, multi-dimensional time-series features (mean, variance, peak value, rising edge slope, falling edge slope, short-time energy, etc.) are extracted. The feature vectors of similar events are then clustered and averaged to form template vectors representing the "typical triggering moment" of that type of event. The template vectors of all categories are aggregated to form the "event trigger template library." This template library serves as prior knowledge for the system, supporting hot updates and continuous iteration. This event trigger template library provides a time-series pattern matching mechanism for subsequent online identification of process events.

[0025] In the online application phase: After loading the pre-built event trigger template library, the system applies a sliding window to the sequence of key sensor data (such as pressure and displacement) acquired in real time. The window length L needs to be set considering the event duration and sampling frequency. For example, for a typical spray gun lifting and lowering action that takes 1-2 seconds, at a sampling frequency of 1Hz, L can be set to 3-5 sampling points. The window sliding step is 1 sampling point. Within each window, a set of multi-dimensional temporal feature vectors consistent with those in the offline phase is extracted.

[0026] The extracted window feature vectors are matched one by one with each template in the event trigger template library for similarity. The matching algorithm adopts Dynamic Time Warping (DTW), which allows two time series to be matched after nonlinear distortion on the time axis, and can effectively overcome the problem of duration scaling of the same event waveform caused by slight differences in operation speed.

[0027] An event is determined to have occurred when the DTW distance between the window feature vector and an event template (such as "backflush start") is less than a preset threshold. This threshold is determined by performing a grid search on historical data during the offline construction phase, with the goal of maximizing the F1-score for event recognition. For example, it can be set to 0.5 times the average DTW distance of all matching pairs.

[0028] Once an event is determined to have occurred, a process event tag with a timestamp is immediately generated. The tag content must include at least the "event type ID".

[0029] Next, based on the process event tag, event parameter extraction is performed. Within the time window of the event's occurrence, sensor data and process event information are jointly read to quantify the event. For example, for "spray gun lifting," the difference between the start and end times of the action can be extracted from the displacement sensor sequence as the "lifting height," while the target set height is read from the PLC log for verification; for "backflushing," the peak value of pressure surges and the duration of pressure drop back to a stable value can be extracted from the pressure sensor sequence. Finally, a complete process event tag is generated, containing the event type, trigger time, and associated specific quantification parameters.

[0030] Step S3: Remove structured interference based on the interference model; This step is one of the core innovations, enabling the "quantitative stripping" of interference.

[0031] Sub-step S31: Pre-construction of the interference model set; It's crucial to clarify the relationship between the "interference model set" and the "event-triggered template library" in step S2. They are progressive: the event-triggered template library is a lightweight model used to detect whether an event has occurred, emphasizing recognition speed and sensitivity, with relatively simplified feature dimensions. The interference model set, on the other hand, is pre-built to provide a model foundation for subsequently determining the target interference model that matches the process event. Furthermore, the models in the interference model set are complete parameterized models used for quantifying and stripping interference, emphasizing replication accuracy and containing complete waveform morphology descriptions. The interference model set can be viewed as a deep extension of the event templates in terms of parameter dimensions. During system runtime, the event type is first identified through the event templates, and then the corresponding interference model is called for stripping based on the identification results.

[0032] The specific construction process of the interference model set is as follows: Select historical furnace sensor data covering various typical operating conditions. Process experts and data analysts collaborate to manually select interference signal intervals caused by various process events (such as lance raising and lowering, backflushing, and slag discharge) using annotation tools. From these intervals, the program automatically extracts a set of quantifiable interference feature parameters to form a parameter vector. The composition of this parameter vector includes, but is not limited to: Pulse amplitude A: The maximum absolute value of the interference signal deviating from the baseline.

[0033] Duration The duration of the interference from its inception to its near-complete end.

[0034] rising slope The average rate of change of the interference signal from 10% peak value to 90% peak value.

[0035] decay time constant For exponentially decaying interference, the time required for the signal to decay from its peak to 36.8% of its peak value.

[0036] Oscillation main frequency The maximum energy frequency obtained after performing a Fast Fourier Transform (FFT) on the interference interval.

[0037] For each type of process event, a Gaussian mixture model (GMM) is used to fit the probability density of the parameter vector distribution of a large number of historical samples. GMM can flexibly capture the multimodal distribution characteristics of parameters. For example, the "backflushing" operation under different batches of raw materials may require different pressure settings, which is reflected in the mixture of Gaussian distributions of amplitude parameters. After training, each type of event has a parameterized statistical distribution model.

[0038] The most crucial step is to determine the corresponding interference waveform basis function template for each type of event. This serves as a bridge linking statistical parameters to physical waveforms. The selection of the template is based on the understanding of the physical mechanism of that type of event: The physical essence of pressure pulse interference caused by "backflush" is the instantaneous release and attenuation of high-pressure gas. The interference waveform basis function template is modeled as an exponentially decaying oscillation function: ; in, The template for the interference waveform basis function is given by: t is the time variable, A is the pulse amplitude (kPa or MPa), τ is the decay time constant (seconds), f is the oscillation frequency (Hz), and φ is the initial phase (radians). At the event trigger time, u(·) is the unit step function, ensuring that the function is within t< The time is 0.

[0039] The displacement measurement interference caused by "spray gun lifting" is essentially a mechanical overshoot and rapid stabilization process of the position sensor, which can be modeled as a damped step response function using a basis function template. ; in, The value represents the disturbance amplitude (related to the displacement change, in mm). The rise time constant (seconds) is the time constant. Let T be the settling time constant (in seconds), and T be the duration of the spray gun action (in seconds). This is the moment the action begins.

[0040] Finally, based on practical experience, smelting engineers verify and correct the typical range of parameters fitted by GMM to ensure that the model does not contradict common sense about the process, thus forming the final set of interference models.

[0041] Sub-step S32: Real-time removal of interference; During real-time processing, the system receives the parameterized process event label (e.g., "Event Type: Backflush, Parameter: Peak Pressure 180 kPa") output from step S2. The system's stripping module retrieves the target interference model corresponding to "Backflush" from the interference model set.

[0042] Utilize real-time parameters carried in the tag (such as And the mean or online estimation obtained from the GMM model. , ), base function template for this event type Instantiate the event to generate a specific sequence of interference signals that perfectly matches the current event. .

[0043] Finally, perform precise time-domain subtraction: ; For time t, The original signal contains interference from process events. The sequence represents the interference signal, indicating the interference components synthesized from the event labels and the model. This refers to the "clean" signal after removing the structured interference from this specific event. This operation is performed point-by-point within the event's influence range.

[0044] Step S4: Extract core state signals under mechanistic constraints; This step is intended to... The core states, such as molten pool temperature, are robustly extracted to ensure their physical rationality. Taking the reconstruction of molten pool temperature as an example, a two-stage framework of "standard Kalman filtering + mechanism constraint projection" is adopted based on process mechanism parameters to solve the problems of inconsistent dimensions and mathematical implementation difficulties in traditional schemes.

[0045] Phase 1: Standard Kalman filtering; To eliminate the dimensional inconsistency, the state variables are... It is defined as "cumulative heat in the molten pool" (unit: kJ), rather than a direct temperature value.

[0046] Construct the state equation and observation equation. State equation: ; in, The heat accumulated in the molten pool at time k. α represents the accumulated heat of the molten pool at time k-1 (the previous time), and α is the comprehensive thermal efficiency coefficient (dimensionless, which includes the conversion of oxygen combustion calorific value and heat absorption efficiency). The oxygen flow rate is time-varying (Nm³ / h). Let h be the step size. The process noise follows a zero-mean Gaussian distribution. The equation of state includes parameters from the process mechanism (time-varying oxygen flow rate). The state evolution term driven by oxygen describes the physical process by which the heat in the molten pool accumulates linearly with the introduction of oxygen.

[0047] Observation equation: ; in, The measured temperature (°C) is after interference removal. The equivalent heat capacity of the molten pool is (kJ / ℃). To mitigate observation noise, this equation maps the intangible thermal state to a measurable temperature value, maintaining dimensional consistency.

[0048] Based on the state equation and observation equation, Kalman filtering is applied to the interference-removed signal. Using the prediction-update recursive formula of standard Kalman filtering, the posterior state estimate and the corresponding temperature estimate at each time step can be obtained. , Let be the posterior state estimate at time k.

[0049] Phase Two: Mechanistic Consistency Projection; This stage aims to impose physical constraints on the Kalman filter results, ensuring that long-term estimates do not deviate from thermodynamic laws. The mechanistic equation (heat accumulation relationship) characterizing the physical laws of the smelting process is transformed into a regularization term, which is then incorporated into the Kalman filter framework as a concrete implementation of the constraints on the smelting mechanism equation. This stage avoids the mathematical difficulties of directly embedding the global regularization term into the recursive formula, employing an independent post-processing projection method.

[0050] At the current time k, define a sliding window of length M, which covers the most recent M estimates. Based on the data within the window, construct a mechanism consistency constraint.

[0051] Determine the time-varying process parameters (i.e., time-varying oxygen flow rate) in the process mechanism parameters. ), and obtain the previous state estimate obtained in the Kalman filter recursion process. .

[0052] Based on the estimated values ​​of the preceding states and time-varying process parameters, the expected values ​​of the current states are calculated using the mechanistic equations. If the heat accumulation in the molten pool perfectly conforms to the linear relationship described by the equations of state, then: ; in, To calculate the expected heat value from the mechanistic equation, the current state-predicted value is... The thermal state at the start of the sliding window (M time points prior). Indicates the overall thermal efficiency coefficient. The length of the sliding window. Let be the time-varying oxygen flow rate at time i within the window. Let be the time step size for the i-th step.

[0053] The objective of the projection operation is to find a corrected heat estimate. This ensures that the posterior state estimate is close to that of the Kalman filter while also satisfying the aforementioned mechanistic constraints.

[0054] Specifically, solve the following constrained minimization problem: ; ; This problem has a closed-form solution, that is... ; in, This is the corrected heat estimate.

[0055] The mechanism consistency residual is calculated based on the current state estimate (i.e., the posterior state estimate at time k) and the current state expectation. .

[0056] With the objective of minimizing the loss function that includes the mechanistic consistency residual, the core state signal is optimally estimated. After mechanistic consistency projection correction, the corrected core state signal, i.e., the corrected heat estimate, is obtained. Then, based on the equivalent heat capacity of the molten pool, the corrected heat estimate is converted into the final temperature estimate. ;in, This is the core state signal that will be output in the final stage. This represents the equivalent heat capacity of the molten pool.

[0057] It should be noted that the equivalent heat capacity of the molten pool is not a fixed constant. During the smelting process, as cold materials, return materials, and other materials are added successively, the total amount of melt in the molten pool gradually increases, and its equivalent heat capacity also increases accordingly. Therefore, in this embodiment, the equivalent heat capacity of the molten pool is estimated in real time based on the amount of raw materials added in the current furnace to obtain heat capacity parameters that match the current melt volume.

[0058] Specifically, ,in, The mass (kg) of the j-th material that has been added is determined by the raw material ratio and the cumulative amount of material added. Let be the specific heat capacity (kJ / (kg·℃)) of the j-th material, and be a known physical property parameter. Trigger whenever a new batch of material is added. The update calculation will be updated. The values ​​are substituted into the observation equation. The amount of raw materials added comes from the raw material ratio data in the process mechanism parameters obtained in step S1.

[0059] Adaptive adjustment strategy: This strategy is crucial for ensuring the long-term robustness of the method. It involves monitoring the mechanistic consistency residuals within the sliding window. The statistical properties of the sequence can be specifically calculated by analyzing the last N... The moving variance. If this variance continues to exceed a preset threshold... (For example, if the fluctuation exceeds the upper limit of normal fluctuation based on historical data statistics), it is determined that the current operating conditions have changed significantly, and the original thermal efficiency coefficient α is no longer applicable. At this time, a two-stage response is triggered.

[0060] Reduce constraints: Adaptively adjust the weight of the regularization term based on statistical characteristics to immediately reduce the strength of the projection constraints. For example, increase the allowable deviation tolerance in the projection optimization problem, or pause the projection operation for several cycles to make the estimate more reliable for the real-time updates of the Kalman filter.

[0061] Online re-estimation: If the high residual state persists, an online re-estimation process for the key parameter (thermal efficiency coefficient α) in the mechanistic equation is triggered based on statistical characteristics. A minimization problem is solved using the data within the current sliding window as the objective: ; in, The expected value of the mechanism at time i is calculated based on α. Let be the posterior state estimate of the Kalman filter at time i within the window. A new thermal efficiency coefficient adapted to the current operating conditions can be quickly solved using a one-dimensional linear search. .

[0062] The online re-estimation process continues until the re-estimated parameters cause the mechanistic consistency residuals to satisfy the preset convergence condition, at which point the updated key parameters are obtained. And update the model.

[0063] Furthermore, if multimodal sensor data needs to be fused, an attention fusion layer can be set before filtering. The feature vectors of the de-interference-reduced signals such as temperature, pressure, and composition are used as input, and a scaled dot product attention module calculates the weights of each modality. A weighted sum is then performed to obtain a comprehensive fused feature that better reflects the core state changes, which is then used as the observation input. Input to the Kalman filter.

[0064] Step S5: Consistency processing of historical data based on process anchor points; This step aims to break down information silos and make historical data usable for model training. Based on preset process anchor points and process mechanism parameters, historical data from different data sources are time-series aligned and feature dimensions are unified.

[0065] Definition and mapping: "Process anchor points" are defined as significant operational moments in the smelting process that have clear physical meaning and are recorded in different systems, such as "start of blowing" (usually defined by the moment the spray gun descends to its position and oxygen supply is turned on) and "completion of slag discharge" (defined by the closure of the slag outlet or the cessation of slag car weighing). A semantic mapping table is constructed for each anchor point, specifying its corresponding field names in different data sources such as DCS locations, MES event records, and laboratory system timestamps. The semantic mapping table serves as a declaration of the mapping relationship between process anchor points and the fields recorded in each data source. As shown in Table 1: Table 1: Semantic Mapping Table (Mapping Relationship Declaration);

[0066] Alignment and Correction: Based on the semantic mapping table, candidate timestamps for each furnace cycle are extracted from historical databases such as DCS and MES, resulting in candidate timestamp sets for each data source. The DCS system time (usually closest to physical reality) is used as the benchmark. For candidate timestamps from other data sources (such as MES), a local search window is used. Inside, among them, The anchor timestamp of the baseline data source, Given the radius of the local search window, the best match is found in the candidate timestamp set. The "local feature distance" of the matching criterion is specified into the following two options: Option A (Multidimensional Feature Matching): Construct a multidimensional feature vector using data from multiple sensors near the anchor point (e.g., temperature, pressure, flow rate). Calculate the Euclidean distance between the multidimensional feature vectors of candidate anchor points in two data sources. The one with the smallest distance is the best match. This method is suitable for data sources with similar sensor configurations.

[0067] Option B (Event Tag Sequence Matching): If both data sources contain event logs, the event tag sequence before and after the candidate anchor point is used as the matching criterion. Event types within a given time period are encoded as strings, and the "contextual similarity" is evaluated by calculating the edit distance between the two strings; the one with the smallest distance is the best match. This method is more reliable because the sequence of event tags directly reflects the process operation context.

[0068] Recommended strategy: Prioritize using Option B (event sequence matching), and degrade to Option A (multidimensional feature matching) when a data source lacks event logs.

[0069] For a series of successfully matched anchor pairs Calculate the local offset: ; in, This is a local offset. For the best match, For the i-th successfully matched anchor pair, the anchor timestamp of the reference data source; These discrete offsets reflect the variation of the clock offset between the two systems along the time axis. Based on the matching results, to obtain the offset at any time throughout the entire furnace cycle, piecewise linear interpolation is performed using these discrete points to form a time offset function that smoothly varies with time. The time offset function is a local offset function obtained by interpolation based on discrete matching points. Finally, point-by-point correction is performed on all timestamps of the MES system: ; in, This is the corrected timestamp for the MES system. The original record timestamp of the MES system. This is a local offset function that changes smoothly over time; By applying different offsets to different time segments using a local offset function, non-constant time bias is corrected; non-constant time bias refers to time deviations with different offsets in different time segments. After this step, the timelines of different data sources are accurately aligned.

[0070] Feature unification and frequency normalization: For missing features, imputation is performed based on the mechanistic correlation of the process stages to obtain complete feature data. For example, if there is no temperature detection in the batching section, a theoretical initial average temperature of the material can be calculated as a temperature feature value based on the measured temperature, specific heat capacity, and environmental heat exchange model of the added material. Finally, using key process anchor points as nodes, the timeline of the entire furnace is divided into multiple intervals (such as the charging period, blowing period, and slag discharge period). Within each interval, an event-based resampling strategy is used to resample data sources with different acquisition frequencies to a unified target frequency (such as 1Hz), resulting in frequency-normalized time-series data. The completed feature data and the frequency-normalized time-series data are combined to form a high-quality process dataset with unified dimensions and time alignment.

[0071] Example 2 Please see Figure 2 This embodiment provides a system architecture corresponding to the method described in Embodiment 1. This system architecture adopts a layered design, comprising, from bottom to top, a multi-source heterogeneous data input layer, a process event semantic perception layer, a semantic interference stripping layer, a smelting mechanism constraint signal reconstruction layer, and a timing alignment and feature completion layer. This architecture drives the entire process through process semantic tags, solving the "signal-semantic separation" problem of traditional methods. The system block diagram can still refer to the architecture diagram of this invention.

[0072] 1. Multi-source heterogeneous data input layer; This layer is used to perform the data acquisition function in step S1 of Embodiment 1. It is configured to collect and aggregate three types of data during the smelting process: Sensor data includes time-series signals such as temperature, pressure, displacement, and composition, sourced from the DCS / PLC system. Temperature data is used for subsequent extraction of core status signals and to assist in determining the intensity of reactions within the furnace; pressure data is used for modeling interference waveforms of process events such as backflushing and for modeling the properties and states of the gas in the spray gun, and can be combined with temperature data to verify furnace conditions; displacement data is used for positioning the spray gun's lifting and lowering movements and for pulse interference removal, and can be used to determine the degree of influence of the incoming gas on the melt morphology; composition data is used to verify whether the reconstructed signal conforms to material balance constraints.

[0073] Process event log: Operation records from the PLC / DCS system, including fields such as event name, trigger time, and associated equipment. This log serves two purposes: firstly, as input to the process event semantic awareness layer for automatic extraction and verification of event tags; and secondly, as detection of "process anchor points," providing a time alignment benchmark for the timing alignment and feature completion layers.

[0074] Process mechanism parameters include time-varying oxygen flow rate, feed ratio, and furnace ID. The time-varying oxygen flow rate serves as the driving variable in the mechanism constraint signal reconstruction layer, and is used to calculate the heat input for the equation of state in conjunction with the characteristics of the feed material and the reaction mechanism. The feed ratio is used to construct material balance constraints and assist in verifying the rationality of the composition data. The furnace ID serves as the primary key for the data index and is used for data association and traceability across processes and systems.

[0075] 2. Semantic perception layer for process events; This layer is used to execute the method in step S2 of Embodiment 1, automatically identifying and classifying process events, and outputting process event labels with timestamps and quantification parameters. This layer adopts an "offline build, online application" deployment paradigm: before deployment, process experts perform offline annotation and feature extraction on historical data to build an event trigger template library; after deployment, this template library is loaded for real-time matching and identification. It internally contains the following sub-modules: Event Trigger Detection Module: Based on sliding window-based temporal pattern matching. A sliding window of length L is set for real-time sequences from sensors such as pressure, flow, and displacement. Within each window, statistical features such as mean, variance, peak value, and rise / fall slope are extracted, or time weights are assigned according to actual production conditions. The extracted feature vectors are then matched against a pre-defined "event trigger template library" using a Dynamic Time Warping (DTW) algorithm. An event is determined to have occurred when the similarity exceeds a pre-defined threshold.

[0076] Event Parameter Extraction Module: Extracts key event parameters from process event logs and sensor sequences. After the event trigger detection module determines that an event has occurred, it jointly extracts key quantitative parameters of the event from the sensor sequences and equipment operation records within that time window, such as spray gun lifting height, backflushing pressure peak value, and duration, for use in the instantiation of the interference model in the subsequent semantic interference stripping layer.

[0077] Semantic interference stripping module: This module executes the method in step S3 of embodiment one, stripping structured interference based on the "process event-interference model library". Its core is a dynamic interference template matching algorithm. This layer contains the following sub-modules: Interference Model Library: This library stores pre-built parameterized interference models corresponding to various process events. This library is progressively related to the "Event Trigger Template Library" in the process event semantic perception layer: the event trigger template library is a lightweight model used to detect whether an event has occurred, emphasizing recognition speed and sensitivity, with relatively simplified feature dimensions; the interference model library contains complete parameterized models used for quantifying and stripping interference, emphasizing replication accuracy and including complete waveform morphology descriptions (e.g., the exponentially decaying oscillation function corresponding to backflushing, and the damped step response function corresponding to lance lifting). During system operation, the event type is first identified by the event trigger template, and then the corresponding model in the interference model library is called for stripping based on the identification result. This model library is constructed using the semi-supervised learning and expert verification method in step S31 of Example 1: historical furnace data is selected, and process events and corresponding interference intervals are manually labeled; signal features (such as pulse amplitude, duration, and rise slope) are extracted from the interference intervals; a Gaussian mixture model (GMM) is used to fit the interference feature distribution of different events; and the model parameters are corrected by smelting engineers based on practical experience.

[0078] Dynamic Interference Template Matching Module: Receives parameterized process event tags from the process event semantic perception layer, retrieves the interference basis function template corresponding to the event type from the interference model library, and instantiates the template using the real-time parameters in the tag to generate a specific interference signal sequence for the current event.

[0079] Interference stripping module: Performs time-domain subtraction operation to strip structured interference components from the original sensor data and outputs a de-interference signal.

[0080] 3. Smelting mechanism-constrained signal reconstruction layer; This layer is used to execute the method of step S4 in Embodiment 1, reconstructing the core state signal under the guidance of the smelting mechanism. To achieve physical consistency and mathematical realizability, this layer adopts a two-stage architecture of "standard Kalman filtering + mechanism consistency projection". It internally contains the following sub-modules: Standard Kalman filter module: Using the noise-reduced temperature measurement as the observation and "accumulated heat in the molten pool" as the state variable, it performs a standard Kalman filter recursive estimation. State equation The thermal efficiency coefficient α and the time-varying oxygen flow rate As a driving term, it describes the linear accumulation process of heat; observation equation Through the equivalent heat capacity of the molten pool This module maps thermal states to temperature values, ensuring dimensional consistency. It outputs the initial posterior thermal estimate and the corresponding temperature estimate for each time step.

[0081] Mechanism Consistency Projection Module: This module imposes physical constraints on the initial posterior estimate output by the standard Kalman filter module. Within a sliding window at the current time, it constructs a heat accumulation mechanism equation driven by the time-varying oxygen flow sequence, calculates the mechanism consistency target value within the window, and projects the initial Kalman filter estimate onto the feasible region defined by this mechanism equation. The projection operation aims to minimize the deviation between the correction and the filter estimate, constrained by satisfying the heat accumulation relationship within the window, and outputs a corrected estimate that satisfies thermodynamic consistency. This module operates as an independent post-processing module, without disrupting the recursive structure of the Kalman filter.

[0082] Adaptive regulator: Monitors the statistical characteristics of the mechanism consistency residuals generated by the mechanism consistency projection module (such as the variance of the residuals within the sliding window), and automatically adjusts the projection constraint strength (such as relaxing the constraint tolerance or pausing the projection operation) when the residuals abnormally increase, making the estimation more reliable for the real-time update of the Kalman filter; when the high residual state persists, it triggers the online re-estimation of the thermal efficiency parameter α in the state equation to adapt to the change in operating conditions.

[0083] Multimodal attention fusion unit (optional): When it is necessary to fuse data from multiple sensors (temperature, pressure, composition, etc.), this unit calculates the weights of the signals after interference removal from each modality through an attention mechanism, performs weighted fusion, and generates a comprehensive feature vector as the observation input of the standard Kalman filter module to highlight the feature representation related to the core state.

[0084] 4. Temporal alignment and feature completion layer; This layer is used to execute the method of step S5 in Embodiment 1, which automatically aligns and completes historical data based on process logic. It contains the following sub-modules: The timeline alignment module uses "furnace ID + process stage" as anchor points and aligns timestamps from different data sources using the Dynamic Time Warping (DTW) algorithm. Its core is based on a pre-built semantic mapping table defining "process anchor points," extracting candidate timestamp sets from each data source, and employing local window association matching to find the best correspondence. The matching criterion prioritizes using the edit distance of event label sequences to assess contextual similarity; when a data source lacks event logs, it degenerates into Euclidean distance matching based on multi-dimensional sensor features. Local offsets are calculated based on discrete matching point pairs, and a locally offset function that smoothly changes over time is generated through piecewise linear interpolation. This function is then used to perform point-by-point correction on the timestamps.

[0085] Feature Dimension Unification Module: This module completes missing features based on process stages. It interpolates missing features from historical data based on the mechanistic relationships between process stages. For example, if there is no temperature sensor in the batching section, a reasonable initial temperature estimate can be calculated based on the initial raw material temperature, ambient temperature, and the thermal balance mechanism equation.

[0086] Frequency normalization module: Employs an event-based resampling strategy. Using key process anchor points as nodes, the timeline of the entire furnace is divided into multiple intervals (such as charging period, blowing period, and slag discharge period). Within each interval, data sources with different acquisition frequencies are resampled to a unified target frequency, generating a high-quality process dataset with unified dimensions and time alignment.

[0087] Through the collaborative work of the above five levels and their internal sub-modules, this system has fully realized the automated processing of the entire chain of "event perception - interference removal - mechanism reconstruction - data consistency", improving the quality of smelting data to a level that meets the needs of advanced applications.

[0088] 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 improving and reconstructing the quality of smelting process data based on process event modeling, characterized in that, The method includes the following steps: Acquire sensor data during the smelting process, process event information reflecting process operation, and process mechanism parameters; Based on sensor data and process event information, identify the currently occurring process events and generate corresponding process event tags; Based on process event labels, target interference models matching the process events are determined from a pre-set set of interference models; By using a target interference model, structured interference is removed from sensor data to obtain a de-interference signal; Based on process mechanism parameters and under the constraints of the smelting mechanism equation, the interference removal signal is reconstructed to extract the core state signal that conforms to physical laws, so as to realize data reconstruction. Based on preset process anchor points, historical data from different data sources are time-series aligned and feature dimensions are unified to generate a consistent process dataset, thereby improving data quality.

2. The method for improving and reconstructing smelting process data quality based on process event modeling according to claim 1, characterized in that, The set of interference models is pre-built in the following way: Acquire historical smelting data and mark the process events and corresponding interference signal ranges; Extract interference characteristic parameters from the interference signal range; By fitting the distribution of interference characteristic parameters corresponding to different types of process events using a statistical model, a parameterized interference model template is obtained. Based on expert experience, the parameters of the interference model template are verified and corrected to form an interference model set.

3. The method for improving and reconstructing smelting process data quality based on process event modeling according to claim 2, characterized in that, Using a target interference model, structured interference is removed from sensor data to obtain a de-interference signal. This process includes the following steps: Obtain the real-time parameters that triggered the current process event; The target interference model is instantiated using real-time parameters to generate the corresponding interference signal sequence; The interference signal sequence is subtracted from the sensor data to obtain the interference-free signal.

4. The method for improving and reconstructing smelting process data quality based on process event modeling according to claim 3, characterized in that, Under the constraints of the smelting mechanism equation, the interference-removing signal is reconstructed to extract the core state signal that conforms to physical laws. The specific steps include the following: The mechanism equations characterizing the physical laws of the smelting process are transformed into regularization terms, and these regularization terms are incorporated into the Kalman filter framework as a specific implementation of the constraints on the smelting mechanism equations. Construct state equations and observation equations, where the state equations include state evolution terms driven by process parameters; Determine the time-varying process parameters in the process mechanism parameters, and obtain the previous state estimate obtained in the Kalman filter recursion process; Based on the estimated value of the previous state and the time-varying process parameters, the expected value of the current state is calculated based on the mechanism equation. Based on the state equation and the observation equation, Kalman filtering is performed on the interference-removed signal to obtain the current state estimate. The mechanism consistency residual is calculated based on the current state estimate and the current state expectation. The core state signal is optimally estimated by minimizing the loss function containing the mechanism consistency residual.

5. The method for improving and reconstructing smelting process data quality based on process event modeling according to claim 4, characterized in that, Mechanism consistency residuals are constructed based on a sliding window to evaluate the deviation between the cumulative change in the state estimate within the window and the expected cumulative change in the mechanism calculated from the time-varying process parameters. The method further includes the following steps: Statistical characteristics of consistency residuals in monitoring mechanisms; The weights of regularization terms are adaptively adjusted based on statistical characteristics to reduce the constraint strength when the mechanism model mismatches; or, the online re-estimation process of key parameters in the mechanism equation is triggered based on statistical characteristics. The online re-estimation process continues until the re-estimated parameters make the mechanism consistency residuals satisfy the preset convergence conditions, and then terminates to obtain the updated key parameters.

6. The method for improving and reconstructing smelting process data quality based on process event modeling according to claim 5, characterized in that, Based on preset process anchor points, historical data from different data sources are time-series aligned and feature dimensions are unified, specifically including the following steps: Define the key operational moments with clear physical significance during the smelting process as process anchor points; Construct a semantic mapping table to establish the association between the record fields corresponding to the same process anchor in different data sources; the semantic mapping table serves as a declaration of the mapping relationship between the process anchor and the record fields of each data source. Based on the semantic mapping table, candidate timestamps for each process anchor point are extracted from each data source to obtain a set of candidate timestamps for each data source. Using the anchor timestamp of one of the data sources as a benchmark, the most matching candidate timestamps in other data sources are found through a local window matching algorithm, and the time offset function that changes with time is calculated accordingly. The timestamps of each data source are corrected point by point using a time offset function to achieve time sequence alignment.

7. The method for improving and reconstructing smelting process data quality based on process event modeling according to claim 6, characterized in that, The time offset function is a local offset function obtained by interpolation based on discrete matching points; The timestamps of each data source are corrected point by point using a time offset function. Specifically, this includes applying different offsets to different time segments using a local offset function to correct non-constant time deviations. Non-constant time deviations refer to time deviations with different offsets in different time segments.

8. The method for improving and reconstructing smelting process data quality based on process event modeling according to claim 7, characterized in that, Based on sensor data and process event information, identify the currently occurring process events and generate corresponding process event tags. This process includes the following steps: A sliding window is applied to the real-time sequence of sensor data, and the temporal features within the window are extracted; A dynamic time warping algorithm is used to calculate the similarity between the temporal features and each template in the preset event triggering template library; When the maximum similarity exceeds a preset threshold, it is determined that the event corresponding to the current template has occurred, and the key quantitative parameters of the event are extracted by combining process event information or sensor data to generate a parameterized process event label.

9. The method for improving and reconstructing smelting process data quality based on process event modeling according to claim 8, characterized in that, Sensor data includes multiple modal data such as temperature, pressure, displacement, and composition; When reconstructing the signal from the interference-reduced signal, the method further includes: The multimodal sensor features are weighted and fused using an attention mechanism to obtain fused features, which enhance the feature representation related to the core state, and the fused features are used as input for signal reconstruction.

10. The method for improving and reconstructing smelting process data quality based on process event modeling according to claim 9, characterized in that, Generating a consistent process dataset includes the following steps: Based on the mechanistic correlation of different process stages, missing features in historical data are filled in to obtain the filled feature data; An event-based resampling strategy is adopted to normalize data sources with different sampling frequencies to a unified time scale, resulting in frequency-normalized time-series data. The completed feature data and frequency-normalized time-series data are combined to form a consistent process dataset.

Images