Prediction system and method for grinding method
By establishing a machine learning model and data conversion formula for the grinding process, the problems of high cost and long time in simulating existing grinding processes have been solved. This has enabled simple and quick prediction of grinding methods, reduced raw material waste and equipment occupation, and improved the stability and production efficiency of the grinding process.
Patent Information
- Application Number
- CN202211736489.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-12-30
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2042-12-30
AI Technical Summary
Existing chemical simulation methods for grinding processes are costly, require numerous input parameters and conditions, and involve long simulation times. They are not suitable for large-scale factory production and grinding processes with variable conditions, resulting in problems such as high failure rates, unstable technology, numerous small-scale trials, long equipment downtime, and raw material waste.
A machine learning model for grinding process is established using a random forest model and an extreme gradient boosting tree model. Combining the data conversion formula and the machine learning model, the prediction process for grinding mode is simplified through a grinding particle size input terminal, a data conversion module, and a grinding mode prediction module. Data fitting is performed using a grinding equipment information interaction module, providing a prediction system and method for grinding mode.
It enables simple and quick acquisition of grinding method data, reduces the number of small-scale tests and equipment occupancy time, reduces raw material waste, ensures the stability of the grinding process and timely delivery, and adapts to the grinding process requirements under varying conditions.
Smart Images

Figure CN118268115B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a prediction system and method for grinding methods. Background Technology
[0002] Due to the stringent particle size control requirements in pharmaceutical formulation manufacturing, grinding technology is typically employed as the final step in the API production process to control particle size. Commonly used grinding technologies include air jet milling and needle-disc milling. The selection of grinding process parameters is largely based on engineers' accumulated experience and judgment. Several sets of grinding parameters are provided for small-scale trials based on the particle size, crystal form, melting point of the uncrushed material, and the acceptable range of particle size after grinding. The results of these trials are then used to determine whether a product meeting the particle size requirements has been obtained. If not, further trials are conducted; if the requirements are met, the appropriate process parameters are selected for scale-up production. This process, based on manual experience, leads to high failure rates, unstable technology, numerous small-scale trials, long equipment downtime, and raw material waste.
[0003] The most commonly used technology on the market is chemical process simulation based on Aspen Plus. However, this tool is expensive, requires a large number of input parameters and conditions, performs well in simulating specific scenarios, and has a long simulation time, making it unsuitable for large-scale factory production and grinding processes with variable conditions. Summary of the Invention
[0004] This invention overcomes the shortcomings of existing chemical simulation methods for grinding processes, such as high cost, numerous input parameters and conditions, and long simulation times. It provides a new prediction system and method for grinding modes. The grinding mode prediction system of this invention can easily and quickly obtain grinding mode data for the material to be ground, at a low cost. It solves the problem of wasting valuable materials caused by uncertainties and limitations of human experience in the grinding process, and facilitates engineers in using recommended parameter and equipment combinations for real-world grinding process operations.
[0005] This invention provides a system for predicting grinding methods, comprising:
[0006] One grinding requirement input terminal is used to input the basic information data of the material to be ground and the target grinding particle size data;
[0007] A grinding particle size prediction module includes a grinding particle size input terminal, a grinding particle size data conversion module, and a grinding particle size prediction output terminal connected in sequence; wherein,
[0008] The grinding particle size input terminal is used to input several propulsion pressures and several grinding pressures of the grinding equipment; the grinding particle size data conversion module is used for data conversion;
[0009] The grinding particle size prediction output terminal is used to output grinding particle size prediction data;
[0010] A grinding mode prediction module includes a grinding demand receiving end, a grinding mode data conversion module, and a grinding demand output end connected in sequence.
[0011] A grinding equipment information interaction module;
[0012] in,
[0013] The grinding demand receiving end is used to receive the target grinding particle size data of the material to be ground;
[0014] The grinding equipment information interaction module is used to receive several propulsion pressure and several grinding pressure data of the grinding equipment, the grinding particle size prediction data, and fit them to the grinding mode data conversion module; the grinding mode data conversion module is used to perform data conversion;
[0015] The grinding demand output terminal is used to output the predicted grinding method data of the material to be ground.
[0016] In this invention, preferably, the grinding particle size data conversion module includes a data conversion formula for data conversion; wherein, obtaining the data conversion formula includes the following steps:
[0017] S11. Collect the grinding process dataset. Each data entry in the grinding process dataset includes the compound ID of the grinding particles, the particle size d90, d50, and d10 before grinding, the feed rate, the pushing pressure, the grinding pressure, the crystal form, the grinding mill equipment information and the grinding energy of the grinding mill, the particle size d90, d50, and d10 after grinding.
[0018] The grinding energy of the grinding mill is obtained by calculation using the following formula:
[0019] Energy = D 2 *N*(P grind +P_atm) / m solid
[0020] D = diameter of the grinding nozzle;
[0021] N = Number of grinding nozzles;
[0022] P grind = Grinding pressure;
[0023] P_atm = 1.013 * 10 5 Pa;
[0024] msolid = Feed rate (g / min);
[0025] S12. A machine learning model for the grinding process is established based on the random forest model and the extreme gradient boosting tree (XGBoost). The grinding process dataset is divided into a training set, a test set, and a validation set. The training set is used to train the model, the test set is used to test the model, and the validation set is used to validate the model. The model trained on the training set is cross-validated to construct the model, thus obtaining the constructed model.
[0026] S13. The grinding process dataset is re-divided into a test set and a training set, and the constructed model is tested on the re-divided test set. 2 Value, calculate the R value of the test 2 The average value;
[0027] S14. Repeat steps S12 and S13 to obtain multiple construction models and the R of the construction models. 2 The average value of R is selected. 2 The model formula with the highest average value.
[0028] In this invention, the grinding particle size data conversion module preferably includes a first database and a grinding particle size data converter. The first database is used to store the data of the grinding process dataset. The grinding particle size data converter is used to construct the data conversion formula based on the data of the grinding process dataset.
[0029] In this invention, the grinding particle size data conversion module preferably further includes a grinding energy data converter, which is used to calculate the grinding energy of the grinding mill according to the calculation formula in step S11 and perform data conversion.
[0030] In this invention, the grinding equipment information interaction module preferably includes a second database and a grinding mode data processor. The second database is used to store several propulsion pressure and several grinding pressure data of the grinding equipment, as well as the grinding particle size prediction data. The grinding mode data processor is used to fit the data based on the data in the second database.
[0031] In this invention, the grinding particle size data conversion module, the grinding equipment information interaction module, and the grinding mode data conversion module are preferably control chips that integrate data conversion and output functions.
[0032] In this invention, the basic information data of the material to be ground preferably includes the planned feed rate, the crystal form of the material to be ground, the particle size d90 before grinding, the particle size d50 before grinding, the particle size d10 before grinding, and the information of the grinding mill equipment.
[0033] In this invention, the target grinding particle size data is preferably the d90 particle size after grinding.
[0034] In this invention, the predicted particle size of the grinding powder is preferably the d90 particle size after grinding.
[0035] In this invention, the input range of the propulsion pressure is preferably 1.5 bar to 10.5 bar; the input range of the grinding pressure is preferably 1 to 10 bar.
[0036] In this invention, the data on several propulsion pressures and several grinding pressures generally exceed 100 sets.
[0037] In this invention, the predicted grinding mode data preferably includes the predicted propulsion pressure and the predicted grinding pressure.
[0038] In this invention, preferably, after step S11, the method further includes screening the grinding process dataset to screen grinding process datasets that are difficult to grind, thereby obtaining a screened grinding process dataset; the screening method preferably satisfies one of the following conditions: Condition 1, the d90 particle size after grinding is greater than 10 micrometers when the pushing pressure is greater than or equal to 9.5 bar; Condition 2, the d90 particle size after grinding is greater than 11 micrometers when the pushing pressure is greater than 8.5 bar.
[0039] In this invention, the data volume of the grinding process dataset is preferably more than 100 records, for example, 148 records. Those skilled in the art will understand that the more data, the better.
[0040] In this invention, the crystal form is preferably the structural information of the crystal form, and the structural information of the crystal form is preferably massive, amorphous, unmeasured, rod-shaped, plate-shaped, or needle-shaped.
[0041] In this invention, the information about the grinding mill preferably includes the brand name, model, container particle size, and number of nozzles in the grinding chamber; wherein, the brand name and model of the grinding mill are preferably DEC MC50 or DEC MC150.
[0042] In this invention, the ratio of the training set, the test set, and the validation set in step S12 is preferably 5:1:1. The ratio of the training set to the test set in step S13 is preferably 5:1.
[0043] In this invention, the cross-validation method is preferably a 50% or 100% method.
[0044] In this invention, the cross-validation preferably includes a grid search method. The hyperparameters are searched using a grid search method. This grid search method can be mainly aimed at the depth of the n_estimators tree (10, 100, step=10), the max_feature (2, 15, step=1), and the max_feature feature method. The hyperparameters obtained by this search method can be a set of hyperparameters with a tree depth of 9, 40 trees, and a max_feature of log2.
[0045] In this invention, the method of re-dividing the test set and training set in step S13 can be conventional in the art, such as random division. The number of divisions can be 50.
[0046] In this invention, the R 2 The best model formula is to generate a .pickle file.
[0047] In this invention, the fitting method can be a conventional fitting method in the art.
[0048] In this invention, the fitted formula can be represented in a graphical manner, for example, as a 3D image with several propulsion pressures, several grinding pressures, and several predicted particle size data of the ground powder as x, y, and z axes, respectively.
[0049] In this invention, those skilled in the art will know that the R 2 It is called the coefficient of determination of the equation, ranging from 0 to 1. The closer it is to 1, the stronger the explanatory power of the equation's variables for y.
[0050] In this invention, those skilled in the art will understand that the random forest model is also called random forest, and the extreme gradient boosting tree is also called XGBoost, which is an implementation of the boosting algorithm. The method based on the random forest model is preferably the regression method using the random forest model. The regression method is also called a regression method. The grid search method is also called a grid search method.
[0051] In this invention, regarding the information about the grinding mill equipment, those skilled in the art will know that when the brand and model of the grinding mill are fixed, the container particle size and the number of nozzles in the grinding chamber are also fixed. The DEC brand MC50 grinding mill equipment has 4 grinding chamber nozzles and a container particle size of 0.95mm. The DEC brand MC150 grinding mill equipment has 6 grinding chamber nozzles and a container particle size of 1.2mm.
[0052] The present invention also provides a method for predicting the grinding mode, which includes the following steps:
[0053] S1. Predict the particle size of the material to be ground.
[0054] S11. Collect the grinding process dataset. Each data entry in the grinding process dataset includes the compound ID of the grinding particles, the particle size d90, d50, and d10 before grinding, the feed rate, the pushing pressure, the grinding pressure, the crystal form, the grinding mill equipment information and the grinding energy of the grinding mill, the particle size d90, d50, and d10 after grinding.
[0055] The grinding energy of the grinding mill is obtained by calculation using the following formula:
[0056] Energy = D 2 *N*(P grind +P_atm) / m solid
[0057] D = diameter of the grinding nozzle;
[0058] N = Number of grinding nozzles;
[0059] P grind = Grinding pressure;
[0060] P_atm = 1.013 * 10 5 Pa;
[0061] m solid = Feed rate (g / min);
[0062] S12. A machine learning model for the grinding process is established based on a random forest model and an extreme gradient boosting tree. The grinding process dataset is divided into a training set, a test set, and a validation set. The training set is used to train the model, the test set is used to test the model, and the validation set is used to validate the model. The model trained on the training set is cross-validated to construct the model, thus obtaining the constructed model.
[0063] S13. The grinding process dataset is re-divided into a test set and a training set, and the constructed model is tested on the re-divided test set. 2 Value, calculate the R value of the test 2 The average value;
[0064] S14. Repeat steps S12 and S13 to obtain multiple construction models and the R of the construction models. 2 The average value of R is selected. 2 The model formula with the highest average value;
[0065] S15. Input several propulsion pressures and several grinding pressures of the grinding equipment to obtain several predicted grinding particle size data.
[0066] S2. Predicted grinding method data for the material to be ground
[0067] S21. A formula for fitting several propulsion pressures and several grinding pressures of the grinding equipment, as well as several predicted particle size data of the grinding equipment.
[0068] S22. Input the target particle size data of the material to be ground to obtain the grinding method data of the material to be ground.
[0069] In this invention, the method for predicting the grinding mode is preferably to use the grinding mode prediction system described above to predict and obtain the grinding mode data of the material to be ground.
[0070] Based on common knowledge in the field, the above-mentioned preferred conditions can be combined arbitrarily to obtain various preferred embodiments of the present invention.
[0071] The reagents and raw materials used in this invention are all commercially available.
[0072] The positive and progressive effects of this invention are as follows:
[0073] 1. The milling mode prediction system of this invention can easily and quickly obtain milling mode data for the material to be milled, avoiding the need for numerous small-scale tests and large quantities of materials in manual design schemes. It reduces mill downtime and the number of central control checks in quality control, uses fewer raw materials, and ensures a high success rate for both R&D milling trials and production, as well as on-time delivery. The milling mode prediction system of this invention will not experience fluctuations in milling performance due to the departure or replacement of milling experts; in actual production, it can meet the demand for predicted milling mode data for 90% of batches of milling.
[0074] 2. The grinding mode prediction method of the present invention uses a machine learning model trained based on data and physical laws to perform simulations with various combinations of conditional parameters. The R-squared value of the model predicting the grinding particle size of the material to be ground is... 2 The average value is 0.862, and the absolute error between the predicted value of the particle size d90 after grinding and the actual value of d90 after grinding is between 0.23 and 1.473 μm. Attached Figure Description
[0075] Figure 1 This is a flowchart of the grinding prediction system used in this invention.
[0076] Figure 2 This is a graph showing basic grinding information displayed by the grinding prediction system of the present invention.
[0077] Figure 3 This is a graph of a grinding plan displayed by the grinding prediction system of the present invention.
[0078] Figure 4 This is another milling plan page diagram displayed by the milling prediction system of the present invention.
[0079] Figure 5 This is a 3D image of Example 1. Detailed Implementation
[0080] The present invention is further illustrated below by way of embodiments, but the invention is not limited to the scope of the embodiments described herein. Experimental methods in the following embodiments that do not specify specific conditions were performed according to conventional methods and conditions, or as selected according to the product instructions.
[0081] Example 1
[0082] This embodiment provides a prediction system for grinding methods, which includes:
[0083] One grinding requirement input terminal is used to input the basic information data of the material to be ground and the target grinding particle size data;
[0084] A grinding particle size prediction module includes a grinding particle size input terminal, a grinding particle size data conversion module, and a grinding particle size prediction output terminal connected in sequence; wherein,
[0085] The grinding particle size input terminal is used to input several feed pressure and several grinding pressure data of the grinding equipment; the grinding particle size data conversion module is used to perform data conversion with a data conversion formula;
[0086] The grinding particle size prediction output terminal is used to output grinding particle size prediction data;
[0087] A grinding mode prediction module includes a grinding demand receiving end, a grinding mode data conversion module, and a grinding demand output end connected in sequence.
[0088] A grinding equipment information interaction module;
[0089] in,
[0090] The grinding demand receiving end is used to receive the target grinding particle size data of the material to be ground.
[0091] The grinding equipment information interaction module is used to receive several propulsion pressure and several grinding pressure data, grinding particle size prediction data of the grinding equipment, and fit the data to obtain data conversion formula 2 to the grinding mode data conversion module; the grinding mode data conversion module is equipped with data conversion formula 2 to perform data conversion;
[0092] The grinding demand output terminal is used to output the predicted grinding method data of the material to be ground;
[0093] The process of obtaining the data transformation formula includes the following steps:
[0094] S11. Collect the grinding process dataset. Each data entry in the grinding process dataset includes the compound ID of the grinding particles, the particle size d90, d50, and d10 before grinding, the feed rate, the pushing pressure, the grinding pressure, the crystal form, the grinding mill equipment information and the grinding energy of the grinding mill, the particle size d90, d50, and d10 after grinding.
[0095] The grinding energy of the grinding mill is calculated using the following formula:
[0096] Energy = D 2 *N*(P grind +P_atm) / m solid
[0097] D = diameter of the grinding nozzle;
[0098] N = Number of grinding nozzles;
[0099] P grind = Grinding pressure;
[0100] P_atm = 1.013 * 10 5 Pa;
[0101] m solid = Feed rate (g / min);
[0102] S12. A machine learning model for the grinding process is established based on the random forest model and the extreme gradient boosting tree. The grinding process dataset is divided into a training set, a test set, and a validation set. The training set is used to train the model, the test set is used to test the model, and the validation set is used to validate the model. The model trained on the training set is cross-validated to build the model and obtain the constructed model.
[0103] S13. The grinding process dataset is re-divided into a test set and a training set, and the constructed model is tested on the re-divided test set. 2 Value, calculate the R value of the test 2 The average value;
[0104] S14. Repeat steps S12 and S13 to obtain multiple construction models and the R of the construction models. 2 The average value of R is selected. 2 The model formula with the highest average value.
[0105] The grinding particle size data conversion module includes a first database and a grinding particle size data converter. The first database is used to store data from the grinding process dataset. The grinding particle size data converter is used to construct a data conversion formula based on the data from the grinding process dataset.
[0106] The grinding particle size data conversion module also includes a grinding energy data converter, which is used to calculate the grinding energy of the grinding mill according to the calculation formula in step S11 and perform data conversion.
[0107] The grinding equipment information interaction module includes a second database and a grinding mode data processor. The second database is used to store several propulsion pressure and several grinding pressure data of the grinding equipment, as well as grinding particle size prediction data. The grinding mode data processor is used to fit the data based on the data in the second database.
[0108] Among them, the grinding particle size data conversion module, the grinding equipment information interaction module, and the grinding mode data conversion module are control chips that integrate data conversion and output functions.
[0109] The basic information data of the material to be ground includes the planned feed rate, the crystal form of the material to be ground, the particle size d90, d50, and d10 before grinding, and the information of the grinding mill.
[0110] The target grinding particle size data is the d90 particle size after grinding.
[0111] Among them, the predicted particle size of the grinding is the d90 particle size after grinding.
[0112] The input range of propulsion pressure is 1.5 bar to 10.5 bar; the input range of grinding pressure is 1 to 10 bar.
[0113] The predicted grinding method data includes the predicted propulsion pressure and the predicted grinding pressure.
[0114] The process includes, after step S11, screening the grinding process dataset, filtering out grinding process datasets that are difficult to grind, and obtaining the filtered grinding process datasets that are easy to grind after excluding them; the screening method is to meet one of the following conditions: Condition 1, the d90 particle size after grinding is greater than 10 micrometers when the pushing pressure is greater than or equal to 9.5 bar; Condition 2, the d90 particle size after grinding is greater than 11 micrometers when the pushing pressure is greater than 8.5 bar.
[0115] In this embodiment, the grinding process dataset contains 148 sets of data, and the filtered grinding process dataset contains 115 sets of data.
[0116] In this embodiment, the propulsion pressure and grinding pressure data are 115 sets.
[0117] Each set of data in the grinding process dataset includes: compound ID of the grinding particles, particle size d90, d50, and d10 before grinding, feed rate, feed pressure, grinding pressure, crystal form, grinding mill equipment information and grinding energy of the grinding mill, particle size d90, d50, and d10 after grinding.
[0118] This grinding process data collection comprehensively covers all crystal types, including blocky, amorphous, undetected, rod-shaped, plate-shaped, and needle-shaped.
[0119] This grinding process data collection includes the brand name and model of the grinding mill, DEC MC50 and DEC MC150. Correspondingly, data on their brand name, model, container particle size, and number of nozzles in the grinding chamber were also collected.
[0120] The data set for this grinding process includes a central propulsion pressure range of 1.5 bar to 10.5 bar and a propulsion pressure range of 1 to 10 bar.
[0121] In step S12, the ratio of training set, test set, and validation set data is 5:1:1.
[0122] In step S13, the ratio of the training set to the test set is 5:1.
[0123] The cross-validation method used was a 50% cross-validation approach.
[0124] Cross-validation includes finding hyperparameters using a grid search method. The grid search method mainly targets the number of trees (10, 100, step=10) of the n_estimators tree, the depth of the max_feature (2, 15, step=1), and the max_feature feature type. The hyperparameters obtained by this method are a set of hyperparameters with a tree depth of 9, 40 trees, and a max_feature value of log2.
[0125] In step S13, the test set and training set are re-divided by randomly dividing the grinding process dataset 50 times according to a ratio of 5:1 between the training set and the test set.
[0126] In step S14, multiple construction models and the R of the construction models are obtained. 2 In the average value, the R of a certain model 2 The average value is the highest, at 0.862. This model is chosen as the data transformation formula. The R-squared value of this model is... 2In step S13, on a randomly partitioned test set, the R of the model... 2 The R-value is 0.963, which is the model's R-value in this randomly partitioned training set. 2 The value is 0.98. This data conversion formula generates a .pickle file.
[0127] The fitting method involves inputting the aforementioned 115 sets of data (propulsion pressure and grinding pressure) to obtain 115 predicted grinding particle size data. Data conversion formula 2 is obtained by fitting the 115 sets of propulsion pressure, grinding pressure, and predicted grinding particle size data, and these are represented by 3D images with x, y, and z axes respectively. The resulting 3D images are shown below. Figure 5 As shown, the 3D image clearly shows the predicted d90 particle size of the material to be ground after grinding under different propulsion and grinding pressures. Figure 5 In the diagram, some points represent propulsion pressure = grinding pressure + 1 bar, while others represent propulsion pressure greater than grinding pressure.
[0128] The basic information diagram of the grinding compound in the grinding mode prediction system of this embodiment 1 is shown below. Figure 2 As shown in the figure, the recommended page for grinding simulation parameters is as follows: Figure 3 As shown in the figure, the recommended display page for grinding simulation parameters is as follows: Figure 4 As shown. Explanation: Figure 2 The information regarding the grinding type and the melting point of the compound is unrelated to the grinding mode prediction data of the grinding mode prediction system of this invention. This embodiment also provides a method for predicting the grinding mode, which uses the grinding mode prediction system of Embodiment 1 described above to predict the recommended feed pressure and recommended grinding pressure of the material to be ground.
[0129] Application Examples 1-3
[0130] In this application, Examples 1-3 use the prediction system of Example 1 to predict recommended grinding methods. The particle size d90, d50, and d10 of the material to be ground before grinding, the planned feed rate, and the crystal form information of Examples 1-3 are shown in Table 1 below. An MC50 grinding mill was used, with target grinding particle sizes D90 of 8 μm, 7 μm, and 6 μm, respectively. These values were input into the prediction system of Example 1, and the recommended feed pressure and recommended grinding pressure data were obtained. The recommended data results for the material to be ground are shown in Table 2 below.
[0131] Table 1. Input information for the materials to be ground in Application Examples 1-3
[0132]
[0133] Table 2. Recommended data results for the materials to be ground in Examples 1-3
[0134]
[0135] Application Examples 4-5
[0136] Application Examples 4-5 use the prediction system of Example 1 to predict recommended grinding methods. The particle size d90, d50, and d10 of the material to be ground before grinding, the planned feed rate, and the crystal form information are shown in Table 3 below. An MC50 grinding mill was used; the target grinding particle sizes D90 were 8 μm and 7 μm, respectively. These values were input into the prediction system of Example 1, and the recommended feed pressure and recommended grinding pressure data were obtained. The recommended data results for the material to be ground are shown in Table 4 below.
[0137] Table 3. Input information for the materials to be ground in Application Examples 4-5
[0138]
[0139] Table 4. Recommended data results for the materials to be ground in Application Examples 4-5
[0140]
[0141] Effect Examples 1-3
[0142] The predictive system for recommending grinding methods is applied in actual production; its flowchart is shown below. Figure 1 .
[0143] The predicted grinding method data of the material to be ground obtained from Examples 1-3 were used for a small-scale workshop test to obtain the actual particle size d90 data after grinding.
[0144] As shown in Table 5 below, the absolute error between the actual d90 particle size data after grinding and the predicted d90 data after grinding obtained using the model of Example 1 is between 0.23 and 0.3 μm.
[0145] Table 5. Results of particle size d90 after grinding in Examples 1-3
[0146]
[0147] Example 4-5
[0148] The predictive system for recommending grinding methods is applied in actual production; its flowchart is shown below. Figure 1 .
[0149] The predicted grinding method data of the material to be ground obtained from Examples 4-5 were used for a small-scale workshop test to obtain the actual particle size d90 data after grinding.
[0150] As shown in Table 6 below, the absolute error between the actual d90 particle size data after grinding and the predicted d90 data after grinding obtained by using the model of Example 1 is between 0.871 and 1.473 μm.
[0151] Table 6. Results of particle size d90 after grinding in Examples 4-5
[0152]
[0153]
[0154] As demonstrated by the recommended propulsion pressure and grinding pressure in the above-described effect embodiments, and through small-scale trials in the workshop, the milling mode prediction system of this invention can easily and quickly obtain milling mode data for the material to be milled. This avoids the need for numerous small-scale trials and large quantities of materials in manual design schemes, reduces mill occupancy time and the number of central control inspections in quality control, uses fewer raw materials and reagents, ensures a high success rate for both R&D milling trials and production, and guarantees on-time delivery. Furthermore, the milling mode prediction system of this invention will not experience fluctuations in milling performance due to the departure or replacement of milling experts; in actual production, it can meet the demand for predicted milling mode data for 90% of batches of milling.
Claims
1. A prediction system for a grinding method, characterized in that, It includes: One grinding requirement input terminal is used to input the basic information data of the material to be ground and the target grinding particle size data; A grinding particle size prediction module includes a grinding particle size input terminal, a grinding particle size data conversion module, and a grinding particle size prediction output terminal connected in sequence; wherein, The grinding particle size input terminal is used to input several propulsion pressures and several grinding pressures of the grinding equipment; the grinding particle size data conversion module is used for data conversion; the grinding particle size prediction output terminal is used to output grinding particle size prediction data. The grinding particle size data conversion module includes a data conversion formula for data conversion. The process of obtaining the data conversion formula includes the following steps: S11. Collect the grinding process dataset. Each data entry in the grinding process dataset includes the compound ID of the grinding particles, the particle size d90, d50, and d10 before grinding, the feed rate, the pushing pressure, the grinding pressure, the crystal form, the grinding mill equipment information and the grinding energy of the grinding mill, the particle size d90, d50, and d10 after grinding. The grinding energy of the grinding mill is obtained by calculation using the following formula: D = diameter of the grinding nozzle, in mm; N = Number of grinding nozzles; = Grinding pressure, the unit of which is Pa; = 1.013 ; = Feed rate, its unit is g / min; S12. A machine learning model for the grinding process is established based on the Random Forest model and the XGBoost tree model. The grinding process dataset is divided into a training set, a test set, and a validation set. The training set is used to train the model, the test set is used to test the model, and the validation set is used to validate the model. The model trained on the training set is cross-validated to construct the model, thus obtaining the constructed model. S13. The grinding process dataset is re-divided into a test set and a training set, and the constructed model is tested on the re-divided test set. 2 Value, calculate the R value of the test 2 The average value; S14. Repeat steps S12 and S13 to obtain multiple construction models and the R of the construction models. 2 The average value of R is selected. 2 The model formula with the highest average value; A grinding mode prediction module includes a grinding demand receiving end, a grinding mode data conversion module, and a grinding demand output end connected in sequence. A grinding equipment information interaction module; in, The grinding demand receiving end is used to receive the target grinding particle size data of the material to be ground; The grinding equipment information interaction module is used to receive several propulsion pressure and several grinding pressure data of the grinding equipment, the grinding particle size prediction data, and fit them to the grinding mode data conversion module. The grinding method data conversion module is used to perform data conversion; The grinding demand output terminal is used to output the predicted grinding method data of the material to be ground.
2. The prediction system for grinding methods as described in claim 1, characterized in that, The grinding particle size data conversion module includes a first database and a grinding particle size data converter. The first database is used to store data of the grinding process dataset. The grinding particle size data converter is used to construct a data conversion formula based on the data of the grinding process dataset. The grinding particle size data conversion module also includes a grinding energy data converter, which is used to calculate the grinding energy of the grinding mill according to the calculation formula in step S11 and perform data conversion.
3. The prediction system for grinding methods as described in claim 1, characterized in that, The grinding equipment information interaction module includes a second database and a grinding mode data processor. The second database is used to store several propulsion pressure and several grinding pressure data of the grinding equipment, as well as the grinding particle size prediction data. The grinding mode data processor is used to fit the data based on the data in the second database.
4. The prediction system for grinding methods as described in claim 1, characterized in that, The grinding particle size data conversion module, the grinding equipment information interaction module, and the grinding method data conversion module are control chips that integrate data conversion and output functions.
5. The prediction system for grinding methods as described in claim 1, characterized in that, The basic information data of the material to be ground includes the planned feed rate, the crystal form of the material to be ground, the particle size d90 before grinding, the particle size d50 before grinding, the particle size d10 before grinding, and the information of the grinding mill equipment. And / or, the target grinding particle size data is the d90 particle size after grinding; And / or, the predicted particle size data is the d90 particle size after grinding; And / or, the input range of the propulsion pressure is 1.5 bar - 10.5 bar; the input range of the grinding pressure is 1-10 bar; And / or, the predicted grinding mode data includes the predicted propulsion pressure and the predicted grinding pressure.
6. The prediction system for grinding methods as described in claim 1, characterized in that, After step S11, the method further includes screening the grinding process dataset to identify grinding process datasets that are difficult to grind, and then excluding them to obtain the filtered grinding process datasets that are easy to grind.
7. The prediction system for grinding methods as described in claim 6, characterized in that, The screening method is to meet one of the following conditions: Condition 1, the d90 particle size after grinding is greater than 10 micrometers when the pushing pressure is greater than or equal to 9.5 bar; Condition 2, the d90 particle size after grinding is greater than 11 micrometers when the pushing pressure is greater than 8.5 bar.
8. The prediction system for grinding methods as described in claim 1, characterized in that, The data volume of the grinding process dataset is more than 100 sets; And / or, the data on the plurality of propulsion pressures and the plurality of grinding pressures are more than 100 sets; And / or, the crystal form is the structural information of the crystal form, and the structural information of the crystal form is massive, amorphous, unmeasured, rod-shaped, plate-shaped or needle-shaped; And / or, the grinding mill equipment information includes the brand name, model, container particle size, and number of nozzles in the grinding chamber of the grinding mill.
9. The prediction system for grinding methods as described in claim 8, characterized in that, The data volume of the grinding process dataset is 148 sets; And / or, the brand name and model of the grinding mill are DEC MC50 or DEC MC150.
10. The prediction system for grinding methods as described in claim 1, characterized in that, The ratio of the training set, the test set, and the validation set in step S12 is 5:1:1; And / or, the ratio of the training set to the test set in step S13 is 5:1; And / or, the cross-validation method is a 50% or 100% method; And / or, the cross-validation includes finding hyperparameters using a grid search method; And / or, the R 2 The model formula with the highest average value is used to generate a .pickle file.
11. A method for predicting grinding methods, characterized in that, It includes the following steps: S1. Predict the particle size of the material to be ground; S11. Collect the grinding process dataset. Each data entry in the grinding process dataset includes the compound ID of the grinding particles, the particle size d90, d50, and d10 before grinding, the feed rate, the pushing pressure, the grinding pressure, the crystal form, the grinding mill equipment information and the grinding energy of the grinding mill, the particle size d90, d50, and d10 after grinding. The grinding energy of the grinding mill is obtained by calculation using the following formula: D = diameter of the grinding nozzle, in mm; N = Number of grinding nozzles; = Grinding pressure, the unit of which is Pa; = 1.013 ; = Feed rate, its unit is g / min; S12. A machine learning model for the grinding process is established based on a random forest model and an extreme gradient boosting tree. The grinding process dataset is divided into a training set, a test set, and a validation set. The training set is used to train the model, the test set is used to test the model, and the validation set is used to validate the model. The model trained on the training set is cross-validated to construct the model, thus obtaining the constructed model. S13. The grinding process dataset is re-divided into a test set and a training set, and the constructed model is tested on the re-divided test set. 2 Value, calculate the R value of the test 2 The average value; S14. Repeat steps S12 and S13 to obtain multiple construction models and the R of the construction models. 2 The average value of R is selected. 2 The model formula with the highest average value; S15. Input several propulsion pressures and several grinding pressures of the grinding equipment to obtain several predicted grinding particle size data. S2. Predict the grinding method data for the material to be ground; S21. A formula for fitting several propulsion pressures and several grinding pressures of the grinding equipment, as well as several predicted particle size data of the grinding equipment. S22. Input the target grinding particle size data of the material to be ground to obtain the grinding method data of the material to be ground; The method for predicting the grinding mode is to use the grinding mode prediction system as described in any one of claims 1-10 to predict and obtain the grinding mode data of the material to be ground.
Citation Information
Patent Citations
Learning model generation method, computer program, learning model, control device, and control method
WO2020245915A1
