A model for predicting the properties of cement clinker based on deep learning and its application

Abstract

A model for predicting the properties of cement clinker based on deep learning and its application. Relating to the fields of materials science and computer science. Comprising the following steps: S1. Data preparation: collecting XRD diffraction data from a large number of cement clinker samples, including scan angles and corresponding peak intensities, and simultaneously collecting the strengths of the cement clinker samples; S2. Data preprocessing: normalizing the XRD diffraction data and then transforming it into a two-dimensional matrix; S3.Model construction: For the transformed and processed data on the composition and strength properties of the samples, a deep neural network model is adopted for learning and training, and their weights are calculated by vector multiplication (product of points) to build a model for predicting the properties of the cement. In step S1, the specific surface area of ​​the cement clinker sample is controlled to ±50 m² / kg. The invention enables rapid prediction of the 28-day compressive strength of the clinker, thus significantly reducing the time required by traditional experimental testing methods.

Description

Title of the invention: A model for predicting the properties of cement clinker based on deep learning and its application Technical field

[0001] The present invention relates to the fields of materials science and computer science, and in particular to a model for predicting the properties of cement clinker based on deep learning and its application. TECHNOLOGICAL BACKGROUND

[0002] Cement is a basic material of the construction industry in our country, its properties directly determine the quality and service life of engineering construction. Cement clinker serves as a key component of cement, its gelling properties not only directly affect the amount of mixing material, but also determine the development of properties of final cement concrete. However, as the most critical mechanical property indicator of cement clinker, the 28-day compressive strength is also the main measurement basis of cement class in the current national standard of Portland cement, and it takes at least 28 days to measure and determine it.This long-cycle and large time lag detection method not only reduces the efficiency of engineering production and application, increases the uncertainty and risk in engineering construction, but also makes it more difficult to control the quality of raw materials in engineering construction.

[0003] Domestic and foreign researchers have conducted a lot of research on methods for rapid determination of cement clinker properties, but no reliable results have been obtained. Since the 1980s, the field of materials science has focused great attention on predicting cement properties by mathematical methods, and many studies have used mathematical statistics to establish empirical formula methods to predict cement clinker properties. For example, the traditional statistical method of unitary or multiple linear regression has been adopted to simplify the nonlinear relationship between cement properties and each detection value into a linear functional relationship.However, due to the complexity and multidimensionality of factors affecting clinker properties, as well as the limitation of data quantity and analysis techniques, it was difficult to obtain ideal results. With the arrival of the fourth paradigm information age, machine learning methods are rapidly emerging, promoting . the change in the research and development mode of materials from "experience + trial and error" to a "data-driven" mode. Existing research uses machine learning methods to establish a complex multidimensional model between material composition, process, and performance, which can accurately predict the composition, structure, organization, and performance of the material. However, because the model is highly dependent on the stability of process parameters, it tends to have poor generalization ability, resulting in the prediction of properties severely deviating from reality when the unknown process parameters fluctuate or even when the working conditions change completely.

[0004] Description of the invention:

[0005] The invention provides a model for predicting cement clinker properties based on deep learning and its application, which adopts raw signal data, such as XRD diffraction data, which contains information on the composition and structure of the material, to transform and use them as input data to construct a cement clinker property prediction model and realize rapid and accurate prediction of cement clinker properties.

[0006] The technical plan of the invention is as follows: a model for predicting the properties of cement clinker based on deep learning, comprising the following steps:

[0007] SL Data preparation: collecting XRD diffraction data of a large number of cement clinker samples, including the scanning angles and corresponding peak intensities, and, at the same time, collecting the resistances of the cement clinker samples;

[0008] S2. Data preprocessing: normalize XRD diffraction data and then transform into a two-dimensional matrix form;

[0009] S3. Model construction: for the transformed and processed data on the composition and strength properties of the samples, a deep neural network model is adopted for their learning and training, and their weight calculation is performed by vector multiplication (point product), so as to build a cement property prediction model. In the SI stage, the specific surface area of ​​said cement clinker sample is controlled at ±50m2 / kg;

[0010] The XRD diffractometer is calibrated using an external standard, the peaks of the scanning angle being between ±0.0025° and ±0.0125° of its standard value, and said external standard is aluminum oxide and silicon oxide.

[0011] In step S1, for data with less than 500 sample groups, it is required that the strongest peak of the XRD diffraction peaks of the tested samples is not less than 10,000 counting points.

[0012] A scanning speed of 0.01~0.02° / step and 0.01s / step~100s / step is adopted to perform XRD test of the samples.

[0013] In step S2, the collected data is cleaned and preprocessed to eliminate data containing missing values ​​or obvious abnormal values;

[0014] At the same time, the XRD diffraction data are normalized with the intensity of the strongest diffraction peak as 100, and the data are segmented and convolutionally transformed to form a matrix.

[0015] In step S3, the deep neural network model is a convolutional neural network (CNN) or a recurrent neural network (RNN).

[0016] It also comprises step S4, Evaluation of the model: carrying out the evaluation of the predictive model using evaluation indicators;

[0017] In step S4, said evaluation indicators are the mean absolute percentage error (MAPE) and the root mean square error (RMSE).

[0018] The prediction requirements are met when the mean absolute percentage error is < 1.5 and the root mean square error is < 1.5.

[0019] Step SI further comprises collecting the particle size and volume in the range of the cement clinker sample

[0020] Step S2 further consists of adding the particle size information to the volume in the range.

[0021] Applying a model to predict cement clinker properties based on deep learning, applying a well-trained prediction model to a real prediction task, by inputting XRD diffraction data of an unknown clinker sample to predict the properties of cement clinker.

[0022] The invention combines XRD diffraction data to construct a cement property prediction model, and applies the developed model to achieve rapid and accurate prediction of cement properties.

[0023] The invention has the following advantages:

[0024] 1. Fast prediction: the use of deep learning model allows to achieve rapid prediction of clinker's 28-day compressive strength, thereby significantly reducing the time required for traditional experimental testing methods.

[0025] 2. High precision: by building a deep learning model based on the data of the parameters of the intrinsic properties of the material, it is possible to make an accurate prediction of the strength of the clinker and reduce the prediction error.

[0026] 3. Cost Reduction: This approach can reduce materials, equipment and time required compared to traditional physical testing, thus reducing costs.

[0027] 4. Wide applicability: This method can be applied not only to specific materials or products, but it can also be extended to predict the strength of other types of materials or products, which offers strong broad applicability.

[0028] 5. Integration of more features: in addition to the data from the technique XRD diffraction, more additional features, such as fineness, can be added, which provides a possibility to obtain more comprehensive and accurate strength predictions.

[0029] 6. Automation and digitalization: This method provides technical support to realize the automation and digitalization of clinker strength prediction, which is in line with the trends of modern intelligent manufacturing and Industry 4.0. Brief description of the drawings

[0030] [Fig.l] shows a flowchart of the invention;

[0031] [Fig.2] shows a diffraction delineation diagram of the two groups of samples at the application end of Example 3. Detailed description

[0032] The invention is, as illustrated in [Fig.l], a model for predicting the properties of cement clinker based on deep learning, the specific execution steps being as follows:

[0033] SL Data preparation: Collect XRD diffraction data of a large number of cement clinker samples, including the scanning angles (20) and the corresponding peak intensities (I), among others, for data with less than 500 sample groups, it is required that the strongest peak of the XRD diffraction peaks of the tested samples is not less than 10,000 counting points. At the same time, collect the measured performances of these clinker samples to serve as the objective function data.

[0034] Collect a large quantity of cement clinker powder samples (e.g. Portland cement clinker) whose specific surface area is controlled to ±50m2 / kg-

[0035] The XRD diffractometer is calibrated using external standards such as alumina and silicon oxide, and the position of peak 20 (Cu Ka) is between ±0.0025° and ±0.0125° of its standard value.

[0036] Unification of standards and uniformity of variables in the experimental environment are ensured by controlling the adjustment of the specific surface area and the calibration of the diffractometer.

[0037] A scanning speed of 0.01~0.02° / step and 0.01s / step~100s / step is adopted to perform the XRD test of the samples, the residence time of each step being such that the strongest counting point of the diffraction peaks of the XRD atlas is not less than 10,000 counting points.

[0038] The invention requires that, for data with fewer than 500 sample groups, the strongest peak of the XRD diffraction peaks is not less than 10,000 counting points. Thus: 1. Improving the signal-to-noise ratio; ensuring a sufficient number of counting points, thereby raising the signal-to-noise ratio, so as to obtain a clearer and more reliable diffraction atlas. 2. Data uniformity; by setting a lower limit of counting points, it is possible to ensure that data are comparable between different samples, especially during quantitative analyses. 3. Avoiding errors; high counting points can reduce errors introduced by instrumentation, sample preparation, or other experimental conditions. 4.Optimize data analysis; a sufficient number of counting points can help determine the position, width, and intensity of peaks more accurately during subsequent phases of data processing and analysis, thereby improving the accuracy of the analysis.

[0039] These requirements are intended to ensure that the XRD data collected is of high quality and suitable for in-depth analysis and research.

[0040] At the same time, the 28-day strength performance data of the clinker samples are tested or collected, so as to obtain data on the scanning angle (20) and peak intensity (I) of the clinker samples, as well as data on the strength performance.

[0041] S2. Data preprocessing: When the number of samples is less than 1000, XRD diffraction data are normalized and then transformed into a two-dimensional matrix form. The main purpose of normalization: to eliminate the influence of data from different sources due to differences in measurement or instrumentation and test conditions. When the number of samples is greater than 1000, normalization processing may not be performed.

[0042] The collected data are cleaned and preprocessed to eliminate data containing missing values ​​or obvious abnormal values. At the same time, the diffraction data at other scanning angles (20) are normalized with the intensity of the strongest diffraction peak as 100, and the data are segmented and transformed by convolution to form a matrix.

[0043] The intensity of the strongest diffraction peak is set to 100 to normalize the data and simplify further data processing. Using 100 as a reference ensures that the intensities of the other peaks are all relative values ​​to the strongest peaks, which helps to compare and identify the characteristics of each peak. The normalization method helps to maintain the uniformity and accuracy of data processing.

[0044] A series of transformations and processing must be performed before data input. The input to the model is a one-dimensional sequence vector and, in order to inject positional information into the sequence, the scan angle data (20) is combined with the corresponding peak intensities (I) to form a matrix. Furthermore, in Example 2, the finesse information is assembled into the original matrix in a similar manner to improve the predictive capability of the model.

[0045] By using a mode of inputting a multidimensional data matrix, the invention not only saves resources and computation time, but also maintains data integrity. If the complete diffraction atlas is directly input, it could cause considerable computations and will take a long time. In addition, inputting only the diffraction information of key mineral phases could ignore some important grain information, which will affect the prediction accuracy.

[0046] Therefore, using this data processing method not only effectively reduces the computational load, but also maintains sufficient data integrity to improve the prediction accuracy.

[0047] S3. Construction of the model:

[0048] A deep neural network model, such as a convolutional neural network (CNN) or a recurrent neural network (RNN), is used to learn and train the transformed and processed data on the relationship between the composition and properties of the samples, in order to build a model for the prediction of cement properties.

[0049] S4. Model Evaluation: The prediction model is evaluated using indicators evaluation such as mean absolute percentage error (MAPE) or root mean square error (RMSE), the smaller the error, the better. Generally, when MAPE is < 1.5 and RMSE is < 1.5, the model can be applied to the actual production process, and this ensures that the model has fairly high accuracy and generalization ability.

[0050] Considering the effect of the fineness of the clinker sample on its strength makes the prediction more accurate. Step S1 further includes collecting the particle size and volume in the range of the cement clinker sample.

[0051] Step S2 further consists of adding the particle size information to the volume in the range, so as to obtain a new matrix.

[0052] Applying a model to predict cement clinker properties based on deep learning, applying a well-trained prediction model to a real prediction task, by inputting XRD diffraction data of an unknown clinker sample to predict the properties of cement clinker.

[0053] In the specific application, Example 1

[0054] More than 600 samples of Portland cement clinker powder from nearly 10 domestic cement plants were collected with a specific surface area controlled at 350m2 / kg. The XRD diffractometer was calibrated using external standards such as alumina and silicon oxide, and the position of peak 20 (Cu Ka) was within 0.01° of its standard value. A scanning speed of 0.01~0.02° / step and 0.0ls / step~ 100s / step was adopted to conduct the XRD test of the samples, with the residence time of each step being such that the strongest counting point of the diffraction peaks in the XRD atlas was not less than 10,000 counting points. At the same time, the 28-day strength performance data of these clinker samples were tested or collected, and the scanning angle (20) and peak intensity (I) data, as well as the strength performance data of 602 groups of clinker samples were obtained.

[0055] The collected data were cleaned and preprocessed to eliminate data containing missing values ​​or obvious abnormal values. At the same time, the XRD diffraction data were normalized with the intensity of the strongest diffraction peak of about 20 = 32.235 as 100, and the data were segmented and transformed by convolution to form a 7285x128 matrix.

[0056] The data was divided into a training set and a test set in a ratio of 7:3, and transform was used to build the cement property prediction model.

[0057] The following elements were included: Projecting the XRD data into a high-dimensional space to improve the model characterization capability and reduce data redundancy through downsampling. Constructing a fully connected graph on the original one-dimensional sequential space, where each angular position is a node. The nodes are interconnected and the relationships between the nodes are represented by weighted edges, the weights are calculated by vector multiplication (dot product) to obtain the adjacency matrix, the global integration information will be compressed into a one-dimensional scalar by a decoder. The decoder contains multiple convolutional layers one-dimensional to gradually compress high-dimensional data to one dimension, and to finally map the output values ​​are mapped in the range of 48 to 72.

[0058] The constructed predictive model was evaluated on the training set. In general, when the model evaluation criterion is that the MAPE is less than 1.5 MPa and the RMSE is less than 1.5 MPa, which indicates that the prediction model has the potential to be applied to real production processes and ensures its high accuracy and generalization ability.

[0059] For the current prediction model, its MAPE value on the training set is 0.00021 MPa and RMSE is 0.0016 MPa. By synthesizing these evaluation indicators, the prediction model meets the standard requirements of production applications and demonstrates excellent prediction capability.

[0060] After a series of trainings, the prediction model has been successfully deployed in a real detection environment. When inputting new clinker sample data, the prediction model is able to output the corresponding clinker strength predictive value. Currently, the prediction error has been controlled within the range of 0.65 MPA, indicating that the prediction model has a relatively high prediction accuracy.

[0061] The prediction performance of the prediction model in practical applications requires continuous monitoring and maintenance. During the operation of the model, one should not only focus on the prediction results, but also compare them with the actual resistance values ​​to evaluate their accuracy.

[0062] In order to further optimize the predictive capability of the model, its prediction results as well as the corresponding actual resistance values ​​will be regularly collected during the actual application. Comparison of these data will help to identify potential deviations of the model and make necessary adjustments to the parameters or structure accordingly. This continuous feedback and optimization mechanism will help to ensure the stability and accuracy of the model prediction in long-term applications.

[0063] The above steps enable rapid and accurate prediction of cement clinker strength, which improves production efficiency and also reduces production costs.

[0064] The detailed description of the invention can be adjusted and optimized according to actual needs, for example by improving the algorithm, expanding the data set, etc. in order to improve the accuracy and practicality of the prediction.

[0065] The invention provides clinker production companies with a fast, accurate, intelligent and variation-independent strength prediction method. process, helping to improve the level of clinker quality control and reduce production costs.

[0066] Example 2

[0067] More than 500 samples of Portland cement clinker powder and their corresponding particle size distributions were collected from nearly seven domestic cement plants. The XRD diffractometer was calibrated using external standards such as alumina and silicon oxide, and the position of peak 20 (Cu Ka) was within 0.01° of its standard value. On the one hand, a scanning speed of 0.01~0.02° / step and 0.01s / step~100s / step was used to conduct XRD tests on the samples, with the residence time of each step being such that the strongest counting point of the diffraction peaks in the XRD atlas was not less than 10,000 counting points. On the other hand, the samples were tested for particle size analysis using Hydro 2000MU(A), using alcohol as dispersant, the analysis range being 0.020-2000 (qm).And the 28-day strength performance data of these clinker samples were tested or collected, and the scanning angle (20) and peak intensity (I) data, particle size (qm) and volume in range (%), and strength performance data were obtained for 712 groups of clinker samples.

[0068] The collected data were cleaned and preprocessed to eliminate data containing missing values ​​or obvious abnormal values. The collected data were divided into a training set and a test set in a ratio of 8:2, and the training set was used to train the prediction model.

[0069] Based on Example 1, a 106-dimensional sequence of particle size distributions was added to connect to the original sequence, and similarly, to incorporate the particle size information into the sequence, the particle size information (qm) was added to the volume in the range (%), so that a matrix of (7285+106)x2 dimensions was obtained. Based on this enhanced dimension, a CNN model was used to construct a cement property prediction model by performing weight calculations by vector multiplication (point product).

[0070] During the model evaluation phase, the root mean square error (RMSE) on the training set was only 0.0005 MPa, which shows that the fitting effect of the model on the training set is remarkable. While on the test set, the model also shows excellent prediction performance with a prediction error of 0.54 MPa, this error being much lower than the standard threshold of 1.5 MPa, indicating that the model has the potential for application in actual production.

[0071] Based on the above evaluation results, the model is ready to be deployed in a real detection environment. When new clinker sample data is input into a real environment, the model is able to quickly output the prediction value of clinker strength, and the prediction error can be controlled within 0.6 MPa, which provides reliable technical support for production decision-making.

[0072] Based on Example 1, the influence of the fineness of the clinker sample on its strength is taken into account, which makes the prediction more accurate. Since fineness serves as an important process parameter, its addition provides richer information to the deep learning model, which may further improve the model's prediction accuracy.

[0073] Example 3

[0074] Under the model constructed based on Example 1, two groups of clinker samples were specially taken from a cement plant for XRD tests and 28-day strength tests to verify the accuracy and robustness of the model. The range of XRD angle 20 scans of these two groups of samples was from 5° to 80° and from 5° to 65°, respectively. The results of the XRD data for the two groups of samples A and B are shown in Figures 2, 2a and 2b, respectively. To simulate a variety of possible cases in practical applications, the input diffraction data were allowed to be continuous fragments and there were no restrictions on their sampling intervals and ranges, but in order to ensure that critical features were captured accurately, each input contained at least a range of critical peak segments with a 20 of 10-55°.

[0075] After the normalized processing of the measured XRD data and the matrix transformation, the direct input was carried out for the prediction of the 28-day strength properties. The prediction results of the 28-day strength properties for the two groups of samples A and B were 57.16 MPA and 58.33 MPa, respectively, and the actual measured results were 57.38 MPA and 58.20 MPa, respectively. The prediction strength errors for the two groups of samples were 0.22 MPA and -0.13 MPa, which were much better than the model evaluation indicators, demonstrating the high accuracy and reliability of the performance model. In addition, this demonstrates the enormous potential of machine learning methods in predicting the processing properties of these complex materials.

[0076] The invention constructs an advanced deep learning model based on the two-dimensional array input by X-ray diffraction to predict the properties of cement clinker. The XRD diffraction atlas covers the basic information about the composition and structure that determine the properties of cement clinker, such as crystal phase composition, crystal lattice parameters, amorphous content, crystallinity and grain size, etc. The raw diffraction data collected by XRD serves as input through modification processing, which avoids situations such as signal loss, caused by various processing processes such as data information extraction, and resulting errors.

[0077] On this basis, the invention not only effectively solves the problem of property prediction caused by process fluctuations or changes in operation mode in existing technology, but also can targetably adjust and control the mixing amount of the mixing material, so as to achieve stable control of cement quality.

Claims

Claims

1. A model for predicting the properties of cement clinker based on deep learning, characterized in that it comprises the following steps: S1. Data preparation: collecting XRD diffraction data of a large number of cement clinker samples, including the scanning angles and corresponding peak intensities, and, at the same time, collecting the resistances of the cement clinker samples; S2. Data preprocessing: normalizing the XRD diffraction data and then transforming them into a two-dimensional matrix form; S3.Model construction: For the transformed and processed data on the composition and strength properties of the samples, a deep neural network model is adopted for their learning and training, and their weight calculation is performed by vector multiplication (point product), so as to construct a prediction model of cement properties.

2. A model for predicting the properties of cement clinker based on deep learning according to claim 1, characterized in that, In step S1, the specific surface area of ​​said cement clinker sample is controlled at ±50m2 / kg; The XRD diffractometer is calibrated using an external standard, the peaks of the scanning angle being between ±0.0025° and ±0.0125° of its standard value, and said external standard is aluminum oxide and silicon oxide.

3. A model for predicting the properties of cement clinker based on deep learning according to claim 1, characterized in that, In step S1, for data with less than 500 sample groups, it is required that the strongest peak of the XRD diffraction peaks of the tested samples is not less than 10,000 counting points.

4. A model for predicting the properties of cement clinker based on deep learning according to claim 3, characterized in that a scanning speed of 0.01~0.02° / step and of O.Ols / step~lOOs / step is adopted to carry out XRD test of samples.

5. A model for predicting the properties of cement clinker based on deep learning according to claim 1, characterized in that, In step S2, the collected data is cleaned and preprocessed to eliminate data containing missing values ​​or obvious abnormal values; At the same time, the XRD diffraction data is normalized with the intensity of the strongest diffraction peak as 100, and the data is segmented and transformed by convolution to form a matrix.

6. A model for predicting the properties of cement clinker based on deep learning according to claim 1, characterized in that, In step S3, the deep neural network model is a convolutional neural network or a recurrent neural network.

7. A model for predicting the properties of cement clinker based on deep learning according to claim 1, characterized in that, It also comprises step S4, Model evaluation: performing the evaluation of the predictive model using evaluation indicators; In step S4, said evaluation indicators are the mean absolute percentage error and the mean square error.

8. A model for predicting cement clinker properties based on deep learning according to claim 7, characterized in that, The prediction requirements are met when the mean absolute percentage error is < 1.5 and the root mean square error is < 1.

5.

9. A model for predicting the properties of cement clinker based on deep learning according to claim 1, characterized in that, Step S1 further comprises collecting the particle size and volume in the range of the cement clinker sample Step S2 further comprises adding the information relating to the particle size to the volume in the range.

10. The application of a model to predict cement clinker properties based on deep learning, characterized in that a well-trained prediction model is applied to a real prediction task, and inputs XRD diffraction data of an unknown clinker sample to predict the properties of cement clinker.

Images