Clinker free calcium prediction method based on causal alignment
By using a causal alignment method, the problems of lag and time delay in f-CaO detection in cement production were solved, achieving high-precision, real-time, and interpretable f-CaO prediction, improving the stability and efficiency of the production process, and reducing energy consumption and model maintenance costs.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-30
- Publication Date
- 2026-03-31
AI Technical Summary
Existing technologies for f-CaO detection in cement production suffer from problems such as detection lag, improper handling of time delay and coupling, coexistence of data noise and slow drift, insufficient causality and interpretability, and weak online and transferability. These issues lead to the inability to adjust production operations in real time and unstable product quality.
By employing a causal alignment-based approach, through multi-source data acquisition and fusion, time alignment of process delay grouping, online filtering with a single-sided Gaussian kernel, causal feature extraction and screening, window statistical sample generation, and prediction model construction, the process mechanism is deeply integrated to achieve high-precision, high-real-time, and highly interpretable f-CaO prediction.
It achieves high-precision f-CaO prediction, ensures real-time adjustment of the production process, improves product quality stability and energy efficiency, reduces model deployment and maintenance costs, and has good interpretability and transferability.
Smart Images
Figure CN121768528A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent control technology for cement production processes. Specifically, this invention is a method for predicting free calcium in clinker based on causal alignment. Background Technology
[0002] During the cement clinker firing process, free calcium oxide (f-CaO) is a key indicator for measuring the degree of clinker reaction and firing quality. Excessive f-CaO content can easily lead to poor cement stability and strength fluctuations, while insufficient f-CaO content usually corresponds to overburning, resulting in increased energy consumption and refractory material consumption. Currently, on-site f-CaO testing in industrial settings mainly relies on manual sampling and chemical analysis in laboratories, with testing cycles typically exceeding 90 minutes. This significant testing lag prevents real-time adjustments to production operations based on current conditions, creating a passive control loop where "quality fluctuations have already occurred before parameter adjustment instructions are issued," making it difficult to guarantee stable and optimized product quality.
[0003] To shorten the time delay between detection and parameter tuning, existing technologies have mainly explored the following paths:
[0004] (1) Increasing the frequency of testing or introducing online analyzers. This method has high equipment investment and maintenance costs, and still cannot completely eliminate the inherent delay from sampling to result output, making it difficult to achieve real-time feedback at the minute level.
[0005] (2) Statistical modeling and traditional soft sensing, such as regression analysis, time series model ARIMA, or mechanistic experience-based models. Although these methods are relatively inexpensive to implement, they generally suffer from poor handling of multiple segments and differentiated time lags in the production process. They typically use a single fixed lag or simple overall alignment, leading to misalignment between variables and targets. At the same time, they lack the ability to integrate multi-source heterogeneous data, and the models have poor portability after changes in raw materials, coal types, or production lines, requiring a large amount of manual recalibration.
[0006] (3) Employ machine learning or deep learning models, such as random forests, gradient boosting trees, and long short-term memory networks. These models possess strong nonlinear fitting capabilities, but if process causal constraints and temporal order are not strictly applied during feature engineering, they are prone to learning spurious correlations or causing future information leakage, resulting in excellent performance in offline evaluations but significantly reduced performance in online applications. Furthermore, black-box models have poor interpretability, making it difficult to diagnose and trace the causes when encountering abnormal operating conditions, and the lack of online filtering and robustness strategies makes them sensitive to data noise and slow drift.
[0007] In summary, existing f-CaO prediction techniques suffer from systemic defects such as detection lag, improper handling of time delay and coupling, coexistence of data noise and slow drift, insufficient causality and interpretability, and weak online and transferability. Summary of the Invention
[0008] The purpose of this invention is to provide a clinker free calcium prediction method based on causal alignment, which can deeply integrate the process mechanism and use the f-CaO prediction method with time sequence and causal constraints as boundary conditions to achieve high-precision, high-real-time performance, strong interpretability and good transferability of online prediction and control.
[0009] To achieve the above objectives, this invention provides a method for predicting free calcium in clinker based on causal alignment, comprising the following steps:
[0010] S1: Multi-source data acquisition and fusion, obtaining process variables from the distributed control system (DCS), obtaining raw material chemical composition and clinker free calcium f-CaO test data from the quality management system (LIMS), and performing timestamp unification and outlier processing;
[0011] S2: Based on time alignment using process time delay grouping, the collected variables are divided into multiple variable groups according to their stages and action paths in the cement production process, and each variable group is assigned a time delay value based on the process mechanism. Based on time delay value The variable data within each variable group are shifted over time to form a unified aligned time axis, and the data is then resampled.
[0012] S3: Single-sided Gaussian kernel online filtering. For each variable data after alignment in step S2, a left-truncated Gaussian kernel is used for causal convolution filtering to eliminate noise and avoid future information leakage, thereby obtaining filtered variable data.
[0013] S4: Causal feature extraction and screening. Based on the process mechanism, a structural causal model is established, and the causal topological relationship and delay between variables are defined. Under the constraints of the causal topological relationship and time sequence, feature variables for predicting free calcium f-CaO in clinker are screened from the filtered variable data obtained in step S3.
[0014] S5: Window statistical sample generation, for the test time of each clinker free calcium f-CaO. Based on the process variable sequence filtered in step S3, at the test time... Aggregate statistics within the upstream causal time window to generate statistical feature samples for model training;
[0015] S6: Based on the feature variables selected in step S4 and the statistical feature samples generated in step S5, construct a prediction model and output the predicted value of free calcium f-CaO in clinker.
[0016] Furthermore, in step S2, the process segment and action path include raw material, preheating, decomposition, calcination, cooling, and inspection stages; and time delay values are assigned to different variable groups. They are not the same.
[0017] Furthermore, in step S2, a continuity verification step is included after resampling to determine the continuity of the resampled time series, and to process the data of non-continuous segments using a self-backfilling or neighborhood backfilling strategy.
[0018] Furthermore, in step S3, the causal convolution formula for one-sided Gaussian kernel online filtering is:
[0019] ;
[0020] in, Let g be the filtered value of the variable group at time t. For the variable values at historical moments, For kernel window length, Parameters for controlling smoothing intensity, The sampling period.
[0021] Furthermore, in step S3, for variables with slow drift characteristics, trend tracking is performed by superimposing single exponential smoothing or Kalman filtering on the basis of single-sided Gaussian kernel online filtering.
[0022] Further, in step S4, causal feature extraction and screening specifically includes: temporal sequence constraint, selecting only historical variables before the prediction time t as candidate features; transit path priority, prioritizing variables according to the causal topological path of "raw material - hot process - cooling - finished product" defined by the structural causal model; statistical dependency screening, using at least one of conditional mutual information, Granger causality test or transit entropy to calculate the temporal dependency strength between candidate features and target variable f-CaO; and redundancy suppression, using minimum redundancy maximum correlation mRMR or Hilbert-Schmidt independence criterion HSIC-Lasso method to remove redundancy from the screened features.
[0023] Further, in step S5, the causal time window is defined as: ;
[0024] in, The time for the f-CaO test is... For window width, The distance is the number of steps. The step size is specified; the aggregated statistical operations performed within the window include calculating at least one of the following: mean, median, standard deviation, maximum, minimum, kurtosis, and skewness.
[0025] Furthermore, in step S5, when generating statistical feature samples, the most recently acquired raw material chemical composition test results and clinker f-CaO test results are combined as auxiliary features.
[0026] Furthermore, in step S6, the prediction model is a regression model based on machine learning and a neural network model based on deep learning.
[0027] The clinker free calcium prediction method based on causal alignment of the present invention has the following advantages:
[0028] (1) The present invention has high prediction accuracy and robustness. By using time delay grouping alignment based on process mechanism and causal feature screening, information leakage is reduced, enabling the prediction model to learn the real causal relationship, thereby improving prediction accuracy and robustness under different working conditions.
[0029] (2) The present invention adopts a single-sided Gaussian kernel online filtering method, which does not require future data, ensuring that the model can perform real-time online reasoning and prediction on the production site, thus meeting the timeliness requirements of closed-loop control.
[0030] (3) The entire process of the present invention is based on the construction of process causal topology, which has a clear temporal relationship, which enhances the interpretability of the prediction results, makes it easier for process personnel to understand and trust, and makes it easy to trace the root cause when anomalies occur.
[0031] (4) The method framework of the present invention has universality. For different production lines, only the time delay parameter needs to be adjusted. With a feature selection list, migration applications can be quickly implemented, greatly reducing the cost of model deployment and maintenance.
[0032] (5) The present invention has significant energy-saving and consumption-reducing benefits. By predicting the changing trend of f-CaO in advance and accurately, it can provide operators or advanced control systems with forward-looking optimized settings, thereby achieving precise adjustment of calcination parameters, effectively reducing energy consumption and raw material waste, and stabilizing and improving product quality. Attached Figure Description
[0033] This manual includes the following figures, which illustrate the following:
[0034] Figure 1 This is an overall flowchart of a clinker free calcium prediction method based on causal alignment according to the present invention; Detailed Implementation
[0035] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings, in order to help those skilled in the art to have a more complete, accurate and in-depth understanding of the inventive concept and technical solution of the present invention, and to facilitate its implementation.
[0036] This invention provides a clinker free calcium prediction method based on causal alignment. Its core lies in constructing an end-to-end prediction framework that balances real-time performance, accuracy, interpretability, and transferability. For example... Figure 1 The system flowchart shown illustrates that this method mainly includes steps such as multi-source data acquisition and fusion, time alignment based on process time delay grouping, online filtering with a one-sided Gaussian kernel, causal feature extraction and screening, window statistical sample generation, and prediction model construction and output. The implementation process of this invention is illustrated below through a specific industrial application scenario.
[0037] This embodiment uses a cement clinker production line as an example. This production line is equipped with a complete distributed control system (DCS), a quality management system (LIMS), and an energy management system.
[0038] S1: Multi-source data acquisition and fusion. First, the required data is collected from various systems in the production line. Process variables are obtained from the Distributed Control System (DCS), and key process parameters are collected at a fixed sampling period ∆t, including but not limited to: feed rate: raw material feed rate, pulverized coal feed rate; air volume: high-temperature fan current, air volume of each chamber fan in the grate cooler; speed: rotary kiln speed, grate speed; temperature: decomposer outlet temperature, rotary kiln head temperature, kiln tail temperature, grate cooler chamber under-grate temperature; pressure: kiln head hood negative pressure, C5 cyclone outlet negative pressure; current: kiln main drive current. These data constitute a high-frequency process variable sequence X(t).
[0039] The raw meal chemical composition R(t) and clinker inspection value Y(t) are obtained from the LIMS quality management system. The raw meal chemical composition R(t) is the percentage value of the raw meal exiting the mill, obtained every hour, such as the lime saturation coefficient KH, silicon content SM, and aluminum content IM. The clinker inspection value Y(t) is the free calcium f-CaO content of the clinker, obtained every hour, i.e., Y(t) = f-CaO(t).
[0040] All collected data are synchronized to the same timestamp on a server. Outliers that clearly exceed the process range are labeled. Based on equipment operating status signals, data points under abnormal operating conditions, such as shutdown or tripping, are removed or marked as missing and not used in subsequent model training.
[0041] S2: Based on time alignment using process time delay grouping, according to the cement process flow of "raw materials—preheating—decomposition—calcination—cooling—inspection", the collected variables are divided into the following variable groups G = {g1, g2, g3} according to the process segment and action path, and each group is assigned a time delay value based on the process mechanism. .
[0042] g1 is the raw material composition group, with a time delay τ1 = 15 minutes. g2 is the decomposition and calcination group, with a time delay τ2 = 5 minutes. g3 is the cooling group, with a time delay τ3 = 2 minutes.
[0043] Then, time shift all data points within each variable group: That is, all data timestamps for raw material component group g1 are shifted forward by 15 minutes, data for decomposition and calcination group g2 by 5 minutes, and data for cooling group g3 by 2 minutes. After shifting, all data are unified onto a pre-aligned timeline. Subsequently, all variables are resampled at 1-minute intervals. For non-equal interval test data, a forward padding method is used, i.e., the most recent valid test value is used to fill the gap until the next test value appears. Then, the resampled sequence is... The process involves checking whether the time interval between adjacent data points deviates from the sampling period ∆t (1 minute) by more than 1 second, and then marking the segment as a "discontinuity segment". For missing values within the "discontinuity segment", a "neighborhood backfilling or self-backfilling" strategy will be adopted in the subsequent step S3, that is, the missing value will be filled by linear interpolation of the nearest valid data before and after the missing time.
[0044] S3: One-sided Gaussian kernel online filtering, for a given set of variables g, for a certain process variable For example, the temperature of the decomposition furnace can be calculated using a one-sided Gaussian kernel filter, and its filtered value at any time t is denoted as... .
[0045] ;
[0046] in, This represents the filtered value of a variable in the variable group g at time t. The window length for the filter kernel is measured in data points. In this example, for rapidly changing temperature and pressure variables, the window length is set to... =10; For relatively slow-changing current variables, set =20. The parameter controls the smoothing intensity to ensure that the data within the window is effectively smoothed. Δt is the data sampling period.
[0047] Taking the temperature of the decomposition furnace as an example, at time t, we need to calculate its filter value. The system will start from time t and go back 9 time points (10 points in total), taking the values at each point. , ,,..., The weighted average is calculated by assigning different weights according to the above formula, and finally the smoothed temperature value at time t is obtained.
[0048] For enhancing the processing of slow-drift variables such as wind turbine current, a single exponential smoothing (SES) can be further superimposed on the Gaussian filtering described above for trend tracking. The update formula is as follows: Where s(t) is the smoothing trend value at time t, and the smoothing coefficient is... Optimization can be achieved automatically through grid search. The final value used for feature engineering is... With trend items The weighted sum or directly As a new feature.
[0049] S4: Establish a structural causal model (SCM):
[0050]
[0051]
[0052]
[0053] For a target f-CaO value Only historical variables prior to time t are considered. Variables along the causal topology defined by the SCM are prioritized, for example, "raw material KH" is prioritized over "kiln current," and "decomposer temperature" is prioritized over "grate cooler pressure." Based on the initial selection, conditional mutual information is used for quantitative screening. The conditional mutual information I(X;Y|Z) between each candidate feature X and the target Y is calculated under a given condition set Z, such as some key parent node variables. Features with high conditional mutual information with Y are selected. Finally, the minimum redundancy maximum correlation (mRMR) algorithm is used to remove redundancy from the selected features. This algorithm finds a subset of features that maximizes the correlation between features and the target while minimizing the redundancy between features. After mRMR screening, a set of approximately 15-20 core features is obtained.
[0054] S5: Transform the time series data into static features usable by the model. For each f-CaO test time t, define a causal window upstream of it. ;
[0055] Set the window width w = 30 minutes, the distance steps d = 1, and the step size s = 1 minute. Therefore, the window is... This refers to the time period from 07:29 to 07:59. For each process variable feature ultimately selected by S4, its filtered sequence within this window is used to calculate a series of aggregate statistics, including: mean, median, standard deviation, maximum, minimum, kurtosis, and skewness. Simultaneously, the most recently acquired raw material chemical composition and clinker f-CaO test value are concatenated as the "prev" feature to provide the model with a longer-term operating context. Finally, all these statistics and the prev feature are concatenated into a sample vector, corresponding to the data at that time. The actual f-CaO test value constitutes a label sample.
[0056] S6: Collect production data from the past three months and generate approximately 2000 valid training samples following steps S1 to S5. A Gradient Boosting Decision Tree (GBDT) model, such as XGBoost or LightGBM, is used for supervised training. The model takes the feature vector generated in S5 as input and outputs the predicted f-CaO value. Whenever a new minute-level timeframe arrives, the system automatically executes steps S1 (collecting new data), S2 (time alignment), and S3 (online filtering). Then, using the feature set determined in S4 and the data within the window defined in S5, a feature vector is generated in real time. This feature vector is input into the trained GBDT model, which immediately outputs a predicted value for the future f-CaO content. This prediction result is displayed in real time to the central control operator, allowing for proactive optimization of key parameters such as rotary kiln speed and coal feed rate, forming a "prediction-optimization" feedforward closed loop to stabilize quality and reduce energy consumption.
[0057] The role and effect of the embodiments
[0058] This invention, based on the concept of causal alignment, divides variables into "raw meal composition group (τ=15min)," "decomposition and calcination group (τ=5min)," and "cooling group (τ=2min)," and performs differentiated time shifts. This refined time-delay grouping and alignment ensures that the input features used for prediction, such as raw meal composition and decomposition furnace temperature, accurately lead the f-CaO results they influence on the timeline, eliminating systematic bias at its source. Compared to existing methods using a single fixed lag, the method in this embodiment more realistically reflects the process mechanism, enabling the prediction model to capture the differentiated impact time sequence of each stage on the final quality, thereby significantly improving the accuracy and mechanistic rationality of the prediction.
[0059] This embodiment employs a left-side Gaussian kernel online filter, applying online filtering to each variable based solely on historical data, and superimposing trend tracking for slowly drifting variables. This method avoids future information leakage, ensures the causal compliance of the model in online applications, and solves the key pain points of inflated offline evaluations and subsequent declines. Furthermore, differentiated filtering strategies, such as setting different windows for rapidly changing temperature and slowly changing current, effectively suppress high-frequency noise and track low-frequency drift simultaneously. This results in a high-quality, smooth, and causally compliant data sequence input to the model, laying a solid foundation for high-precision and robust predictions.
[0060] This embodiment constructs a complete feature engineering chain, from "establishing the SCM model" to "initial selection of time and path constraints" to "statistical screening of conditional mutual information" to "mRMR redundancy suppression." This significantly reduces the probability of spuriously related features being selected into the model, ensuring that the model learns genuine causal relationships rather than data coincidences. Therefore, the model's interpretability is enhanced, allowing process engineers to understand the reasons behind the predictions. Simultaneously, because the features are based on robust causal relationships, the model's generalization ability is significantly improved when facing changes in raw materials and coal types, reducing costs.
[0061] This embodiment provides a complete and detailed implementation blueprint, from data acquisition, preprocessing, alignment, filtering, feature engineering, sample generation to model training and online inference. It offers valuable reference for implementation on other production lines. By standardizing and streamlining the entire process, this invention can be quickly migrated to different production lines, requiring only adjustments to the time delay parameters. The feature selection list is sufficient, which meets the requirements of industrial applications for reproducibility and low maintenance costs.
[0062] The present invention has been described above by way of example with reference to the accompanying drawings. Obviously, the specific implementation of the present invention is not limited to the above-described manner. Any non-substantial improvements made using the inventive concept and technical solution; or the direct application of the inventive concept and technical solution to other situations without modification, are all within the protection scope of the present invention.
Claims
1. A method for prediction of free lime in clinker based on causal alignment, characterized by, The method comprises the following steps: S1: multi-source data acquisition and fusion, obtaining process variables from a distributed control system (DCS), obtaining raw material chemical composition and clinker free calcium f-CaO test data from a quality management system (LIMS), and performing timestamp unification and abnormal value processing; S2: Time alignment based on process time lag grouping, the collected variables are divided into multiple variable groups according to their sections and action paths in the cement production process, and each variable group is given a time lag value based on the process mechanism ; based on the time lag value time-shifting the variable data within each variable group, forming a unified aligned time axis, and resampling the data; S3: unilateral Gaussian kernel online filtering, for each variable data aligned in step S2, a left-truncated Gaussian kernel is used for causal convolution filtering to eliminate noise and avoid future information leakage, and filtered variable data is obtained; S4: causal feature extraction and screening, a structural causal model is established based on process mechanism, the causal topological relationship and delay between variables are defined; under the constraints of the causal topological relationship and time sequence, the filtered variable data obtained in step S3 is screened to obtain characteristic variables for predicting clinker free calcium f-CaO; S5: Window statistical sample generation, for each clinker free calcium f-CaO inspection moment , based on the filtered process variable sequence of step S3, the statistical features sample for model training is generated by aggregated statistics within the causal time window upstream of the inspection moment ; S6: based on the characteristic variables screened in step S4 and the statistical feature samples generated in step S5, a prediction model is constructed, and the prediction value of clinker free calcium f-CaO is output.
2. The method for predicting free lime in clinker based on causal alignment according to claim 1, characterized in that, In step S2, the process and the action path include raw material, preheating, decomposition, sintering, cooling and testing; the time delay value given to different variable groups are different.
3. The clinker free lime prediction method based on causal alignment according to claim 1 or 2, characterized in that, In step S2, after resampling, a continuity checking step is further included, which judges the continuity of the resampled time series, and uses self-backfilling or neighborhood backfilling strategy to process the data of non-continuous segments.
4. The method of claim 1, wherein, In step S3, the causal convolution formula of the unilateral Gaussian kernel online filtering is: ； wherein, is the filtered value of the variable group g at time instant t, is the value of the variable at a historical time instant, is the kernel window length, is a parameter that controls the smoothing strength, is the sampling period.
5. The clinker free lime prediction method based on causal alignment according to claim 4, characterized in that, In step S3, for variables with slow drift characteristics, on the basis of the unilateral Gaussian kernel online filtering, a single exponential smoothing or Kalman filtering is superimposed for trend tracking.
6. The clinker free lime prediction method based on causal alignment according to claim 1, characterized in that, In step S4, the causal feature extraction and screening specifically includes: time sequence constraint, only selecting historical variables before the prediction time t as candidate features; transmission path priority, prioritizing variables according to the causal topological path of "raw material-thermal process-cooling-finished product" defined by the structural causal model; statistical dependency screening, using at least one of conditional mutual information, Granger causality test or transfer entropy to calculate the time sequence dependency strength between candidate features and target variable f-CaO; redundancy suppression, using the minimum redundancy maximum correlation mRMR or Hilbert-Schmidt independence criterion HSIC-Lasso method to remove the redundancy of the selected features.
7. The clinker free lime prediction method based on causal alignment according to claim 1, characterized in that, In step S5, the causal time window is defined as: ; wherein, is the f-CaO test moment, is the window width, is the distance step, is the step size; the aggregate statistical operation performed within the window includes calculating at least one of a mean, a median, a standard deviation, a maximum value, a minimum value, kurtosis, and skewness.
8. The method of claim 1, wherein, In step S5, when generating statistical feature samples, the latest obtained raw material chemical composition test results and clinker f-CaO test results are spliced as auxiliary features.
9. The clinker free lime prediction method based on causal alignment according to claim 1, characterized in that, In step S6, the prediction model is a regression model based on machine learning and a neural network model based on deep learning.