Melt index soft-sensing method, equipment and medium based on classification and data enhancement

Abstract

The present invention relates to a soft measurement method, device and medium of melt index based on classification and data enhancement, which is applied to extrusion equipment or other process industrial scenes. The method comprises the following steps: collecting sensor data of extrusion equipment and powder melt index data of feed port and inputting them into a classification model to obtain a predicted production grade; inputting the powder melt index and sensor data into a prediction model corresponding to the production grade to obtain a predicted pellet melt index of the extrusion equipment discharge port, thereby realizing soft measurement of melt index. Aiming at the problem that the sampling frequency does not match the offline characterization data and affects the prediction accuracy, the present invention calculates the maximum historical data time during the training process, constructs a sample set independent of each other based on the sensor data and powder melt index data within the historical data time length, trains the model based on the sample set, realizes parameter optimization by gradually reducing the historical data time, and overcomes the difficulty of estimating the time parameter during the modeling process.

Description

Technical Field

[0001] The invention relates to the field of melt index soft measurement, and in particular to a melt index soft measurement method, equipment and medium based on classification and data enhancement. Background Art

[0002] Polyolefins are widely used in the automotive, electronics, food and other industries due to their excellent properties. Melt flow index (MFI) is one of the key technical indicators to characterize the performance of polyolefin products, and is often used to monitor and control the intermediate process and product quality. The laboratory method is one of the most common offline MFI measurement methods. Both pellets and powders can be tested, but the time lag is long, which is not conducive to real-time control of product quality. Some factories install MFI online analyzers to improve the timeliness of MFI measurements. However, this not only requires additional equipment investment, but also requires frequent mold replacement and recalibration.

[0003] Soft sensing methods provide strong support for real-time monitoring of key variables by establishing models between difficult-to-measure key variables (such as product quality) and easy-to-measure process variables (such as temperature, pressure, etc.). Soft sensing methods usually do not require or rarely add additional hardware and do not affect production processes. They have been widely favored by the industry in recent years. Soft sensing methods can be divided into "model-driven" and "data-driven". "Model-driven" is mainly based on models built on chemical and physical properties. For polyolefins, it involves the use of differential equations to describe processes such as monomer concentration, material diffusion, and chemical reactions. Although this model is an accurate description of the actual process, due to the complexity of the actual situation, the model is often difficult to establish, solve, or has poor generalization ability. "Data-driven" soft sensing makes full use of historical data accumulated in the production process to build a model to predict key variables in real time, showing great potential and receiving widespread attention from industry and academia.

[0004] The existing soft measurement methods for melt index have the following problems:

[0005] (1) The sampling frequency of the sensor does not match the offline characterization data (i.e., the MFI measured in the laboratory), which affects the accuracy of the model.

[0006] (2) The impact of changes in production grades and production processes on measurement results is not fully considered. Summary of the invention

[0007] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a melt index soft measurement method, device and medium based on classification and data enhancement, so as to solve or partially solve the problem that the sampling frequency does not match the offline characterization data, affecting the prediction accuracy.

[0008] The purpose of the present invention can be achieved by the following technical solutions:

[0009] One aspect of the present invention provides a soft measurement method for melt index based on classification and data enhancement, which is applied to extrusion equipment. The soft measurement method for melt index comprises the following steps:

[0010] Collecting sensor data of the extrusion equipment and powder melt index data of the feed port and inputting them into the classification model to obtain a predicted production grade;

[0011] The powder melt index data and the sensor data are input into the prediction model corresponding to the production grade to obtain the predicted melt index of the pellets at the discharge port of the extrusion equipment, thereby realizing soft measurement of the melt index.

[0012] The training process of the classification model and the prediction model includes the following steps:

[0013] Calculate the maximum historical data time, and construct independent sample sets based on the sensor data and powder melting index within the historical data time length;

[0014] The sample set is divided into multiple intervals based on the production brand, and a prediction model corresponding to each interval is obtained through training;

[0015] Training a classification model based on the sample set;

[0016] By reducing the historical data time, the melt index soft measurement accuracy index under different historical data times is calculated, and the classification model and prediction model corresponding to the historical data time with the best measurement accuracy index are used as the finally trained classification model and prediction model.

[0017] As a preferred technical solution, the process of obtaining the prediction model corresponding to each interval through training includes:

[0018] For each interval, ANN, GBDT and other deep learning regression models are constructed respectively. Based on the sample set data of the current interval, ANN, GBDT and other deep learning regression models are trained respectively, and the model with the best prediction accuracy index is selected as the prediction model corresponding to the current interval.

[0019] As a preferred technical solution, the calculation process of the maximum historical data time includes:

[0020] Calculating the material residence time based on the effective volume of the extrusion device and the material flow rate;

[0021] Based on the material residence time, the maximum historical data time is obtained through normal distribution estimation.

[0022] As a preferred technical solution, the process of constructing independent sample sets based on sensor data and powder melting index within the time length of historical data includes:

[0023] Obtain sensor data time series, powder melt index time series and pellet melt index time series;

[0024] For the pellet melt index time series, screening sub-time series that are independent of each other between data points, wherein the time length between any two adjacent pellet melt index corresponding time points in the sub-time series is greater than or equal to the historical data time;

[0025] For each pellet melt index in the sub-time series, the sensor data points and powder melt index time points within the time length range of the historical data before the time point corresponding to the current pellet melt index are selected to form a sample. This step is repeated to obtain multiple samples to achieve the construction of a sample set.

[0026] As a preferred technical solution, the classification model is an ANN model.

[0027] As a preferred technical solution, the division into multiple intervals based on production grades comprises the following steps:

[0028] The sample set is divided into a plurality of intervals by constructing a pellet melt index distribution histogram of the sample set or the like.

[0029] As a preferred technical solution, in the soft measurement process of the melt index, after obtaining the predicted melt index of the pellets at the discharge port of the extrusion device, the following further comprises:

[0030] Based on the obtained laboratory-measured pellet melt index and the predicted pellet melt index at the outlet of the extrusion device, a corrected pellet melt index is obtained through correction.

[0031] As a preferred technical solution, the sensor data includes the extrusion equipment motor torque and speed, gear pump torque and speed, powder feed amount and peroxide feed amount, extruder barrel temperature, extruder barrel pressure, throttle valve opening, etc.

[0032] Another aspect of the present invention provides an electronic device, comprising: one or more processors and a memory, wherein the memory stores one or more programs, and the one or more programs include instructions for executing the aforementioned classification and data enhancement based melt index soft measurement method.

[0033] Another aspect of the present invention provides a computer-readable storage medium, comprising one or more programs for execution by one or more processors of an electronic device, wherein the one or more programs include instructions for executing the aforementioned classification and data enhancement-based melt index soft measurement method.

[0034] Compared with the prior art, the present invention has at least one of the following beneficial effects:

[0035] (1) Overcoming the problem of mismatch between sampling frequency and offline characterization data: During the training process, the maximum historical data time is calculated, and a sample set independent of each other is constructed based on the sensor data and powder melt index within the historical data time length. The model is trained based on the sample set, and the soft measurement accuracy index of the melt index under different historical data times is calculated by gradually reducing the historical data time. The classification model and prediction model corresponding to the historical data time with the best measurement accuracy index are used as the final trained classification model and prediction model. This method solves the difficulty of estimating time parameters in the modeling process without simulating the time characteristics of the process.

[0036] (2) Avoid mutual interference between samples that affect the training results: In the process of constructing the sample set, the sub-time series between data points that are independent of each other are screened from the pellet melt index time series, and the sensor data points and powder melt index time points within the time length range of the historical data before the corresponding time point of the pellet melt index are selected to form a sample to achieve data enhancement, thereby ensuring the independence of each sample and improving the accuracy of the prediction.

[0037] (3) Taking into account the impact of production brand on prediction results: By setting up a classification model to predict the product brand (i.e., product type), and training the prediction model in different intervals based on the product brand, the reliability is improved while the accuracy of the prediction is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 Flow chart of the soft sensing method of melt index based on classification and data enhancement in the embodiment;

[0039] Figure 2 Schematic diagram of the actual melt index soft measurement method in the embodiment. DETAILED DESCRIPTION

[0040] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0041] Example 1

[0042] In view of the problems existing in the above-mentioned prior art, this embodiment provides a soft measurement method of melt index based on classification and data enhancement, which is applied to extrusion equipment such as extruders. The extrusion equipment processes the powder fed into the inlet into granules and then sends them to the discharge port. Figure 1 , the method comprises the following steps:

[0043] S101, analyzing the physical process to determine the influence of time lag and time constant on MFI.

[0044] By analyzing the volume size of the extruder, the characteristics of the processed materials, and the production process parameters in detail, the residence time of the material in the extrusion equipment is calculated to be 8 minutes according to the following formula.

[0045]

[0046] S102, estimating process characteristics with normal distribution, and calculating the longest impact time (ie, the maximum historical data time) ΔT between input and output to be 25 minutes.

[0047] S103, taking each pellet MFI time point as the base point, preliminarily setting the time length of the historical data of each input variable, and converting the time series data into independent data through data enhancement.

[0048] All sensor data within 25 minutes (including extruder main motor torque and speed, gear pump torque and speed, powder feed amount, peroxide feed amount, etc.), as well as the powder melt index data uploaded from the laboratory at the extruder inlet, are used as the independent variable X, and the laboratory pellet MFI test data at the extruder outlet is used as the dependent variable y. Under the premise of achieving mutual independence of each y value, a data pair is constructed with the corresponding X as a sample. The sample construction process includes:

[0049] Obtain sensor data time series, powder melt index time series and pellet melt index time series;

[0050] For the pellet melt index time series, the sub-time series whose data points are independent of each other are selected. Specifically, in the sub-time series, the time length between any two corresponding time points of the pellet melt index adjacent in time is greater than or equal to the historical data time calculated in S102.

[0051] For each pellet melt index in the sub-time series, select the sensor data points and powder melt index time points within the time length range of the historical data calculated in S102 before the time point corresponding to the current pellet melt index to form a sample. Repeat this step to obtain multiple samples to achieve the construction of the sample set.

[0052] S104, based on the distribution of product MFI in historical data, divide the sample set into several intervals.

[0053] By drawing a pellet melt index distribution histogram or other statistical methods, the sample set in step S103 is divided into four intervals according to the production brand (i.e., the type of product, such as low melt index products, melt index between 1-3; medium melt index products, melt index between 15-17; higher melt index products, melt index between 25-30; high melt index products, melt index above 35, and the specific classification method can be adjusted as needed), and then steps S105 and S106 are executed.

[0054] S105, according to the data characteristics of different intervals, train machine learning or deep learning regression models such as ANN and GBDT to improve prediction accuracy.

[0055] In 4 different data intervals, the sample set of the current interval is further randomly divided into training set, validation set, and test set. By randomly selecting samples from the training set, multiple prediction models are trained separately, including but not limited to ANN, GBDT and other machine learning and deep learning regression models.

[0056] S106, constructing a global ANN classification model for predicting the brand of the current product being produced, calling corresponding regression models on different production brands, and other classification models may also be used.

[0057] S107, combining the results of classification and regression, giving the soft measurement prediction result of MFI at any time, and predicting the MFI of polypropylene on the test set.

[0058] S108, compare the RMSE, R2 and other evaluation indicators of the model, if it is not optimal, return to step S104 for iterative training. If it is optimal, go to step S109.

[0059] S109, adjusting data enhancement parameters to further optimize the prediction results.

[0060] By gradually setting lower ΔT from the highest 25 minutes, observe the model RMSE, R2 and other evaluation indicators, if the optimal is not reached, return to step S102. If the optimal is reached, go to step S110;

[0061] S110, the fusion model, runs online and gives predictions in real time.

[0062] In this embodiment, the melting index interval is divided into two categories, low and high, through the ANN model, and the classification accuracy is high; in the low melting index interval, the ANN model is used, and in the high melting index interval, the GBDT model is used; ΔT is set to 20 minutes, the model is integrated, and the actual operation is put into operation.

[0063] See also Figure 2 Schematic diagram of the actual on-line soft measurement process of melt index, which includes the following steps:

[0064] Step 1, read in data. The data source includes sensor data collected in real time from the DCS system through the OPC server and powder test data from the laboratory.

[0065] Step 2, configure the environment on a common computer, run the trained classification model and prediction model, and calculate the predicted value of the melt index;

[0066] Step 3, read the pellet test data from the laboratory, and calibrate the model prediction result based on the pellet test data and the melt index prediction value to obtain the melt index correction prediction value.

[0067] It should be noted that, in the absence of conflict, this method can also be applied to the soft measurement process of melt index or other difficult-to-measure key variables in other process industrial scenarios other than extrusion equipment.

[0068] In order to verify the effectiveness of this method, the prediction effect of this method is compared with that of other methods. The results are shown in Table 1.

[0069] Table 1 Comparison between this method and other methods

[0070] method RMSE MAE MRE <![CDATA[R 2 ]]> Correlation SVR 1.68 0.956 0.071 0.986 0.992 DTR 1.16 0.554 0.037 0.993 0.996 GBDT 0.626 0.390 0.032 0.997 0.999 GPR 1.136 0.416 0.075 0.993 0.997 ANN 1.28 0.805 0.069 0.992 0.996 This method 0.600 0.383 0.031 0.998 0.999

[0071] This method has the following beneficial effects:

[0072] (1) Convert time series data into independent data. Since the sampling frequency of the sensor does not match the offline characterization data (MFI measured in the laboratory), some methods align the input signal with the output by simulating the time characteristics of the process. However, the accuracy and robustness of this simulation are not ideal. This method is based on the physical mechanism of the processing process. By reasonably selecting the historical data time (i.e., data enhancement parameters), taking each output MFI as a benchmark, all factors that may affect the output MFI are included in the input data set, thereby converting the time series data into independent data. This method overcomes the difficulty of estimating time parameters in the modeling process, reconstructs the time series data through data independence, and improves the accuracy of the model.

[0073] (2) Train a simple global classification model to identify the current production brand, and train regression models for different brands of products. Experiments have shown that it is difficult for a single model to capture the changes in MFI of different brands, and existing methods have designed more complex steady-state and transition state models. Based on the statistical analysis of the MFI distribution, this work does not distinguish between transition states and simply divides the intervals. The best regression models are trained in different intervals for real-time prediction of MFI; a global classification model is trained to indicate the current brand and switch the corresponding regression model. This method improves both the accuracy and reliability of the model.

[0074] (3) Adjust the enhancement parameters to adapt to changes in production process. Considering that the time characteristics of the production process may be different under different working conditions, the upper limit of the data enhancement parameters is first estimated, and then the enhancement parameters are optimized according to the fitting results. The adaptability of the model is further improved by the present invention.

[0075] (4) When running online, the prediction model can be realized based on the factory's existing DCS, OPC and laboratory sampling system data as input, using more mature devices, such as ordinary computers, so the cost is low and the reliability is high.

[0076] Example 2

[0077] This embodiment provides an electronic device, including: one or more processors and a memory, wherein the memory stores one or more programs, and the one or more programs include instructions for executing the melt index soft measurement method based on classification and data enhancement as described in Example 1.

[0078] Example 3

[0079] This embodiment provides a computer-readable storage medium, including one or more programs for execution by one or more processors of an electronic device, wherein the one or more programs include instructions for executing the soft measurement method for melt index based on classification and data enhancement as described in Example 1.

[0080] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A soft-sensing method for melt index based on classification and data enhancement, characterized in that: Applied to extrusion equipment, the melt index soft measurement method comprises the following steps: Collecting sensor data of the extrusion equipment and powder melt index data of the feed port and inputting them into the classification model to obtain a predicted production grade; The powder melt index data and the sensor data are input into the prediction model corresponding to the production grade to obtain the predicted melt index of the pellets at the discharge port of the extrusion equipment, thereby realizing soft measurement of the melt index. The training process of the classification model and the prediction model includes the following steps: Calculate the maximum historical data time ΔT, and construct independent sample sets based on the sensor data and powder melting index within the historical data time length; The sample set is divided into multiple intervals based on the production brand, and a prediction model corresponding to each interval is obtained through training; Training a classification model based on the sample set; By reducing the historical data time, calculating the melt index soft measurement accuracy index under different historical data times, and taking the classification model and prediction model corresponding to the historical data time with the best measurement accuracy index as the final trained classification model and prediction model, The process of obtaining the prediction model corresponding to each interval through training includes: For each interval, we build ANN, GBDT and deep learning regression models respectively. Based on the sample set data of the current interval, we train ANN, GBDT and deep learning regression models respectively, and select the model with the best prediction accuracy index as the prediction model corresponding to the current interval. The process of constructing a sample set independent of each other based on the sensor data and the powder melting index within the time length of the historical data includes: Obtain sensor data time series, powder melt index time series and pellet melt index time series; For the pellet melt index time series, screening sub-time series that are independent of each other between data points, wherein the time length between any two adjacent pellet melt index corresponding time points in the sub-time series is greater than or equal to the maximum historical data time ΔT; For each pellet melt index in the sub-time series, the sensor data points and powder melt index time points within the time length range of the historical data before the time point corresponding to the current pellet melt index are selected to form a sample. This step is repeated to obtain multiple samples to achieve the construction of a sample set.

2. The soft sensing method of melt index based on classification and data enhancement according to claim 1, characterized in that: The calculation process of the maximum historical data time includes: Calculating the material residence time based on the effective volume of the extrusion device and the material flow rate; Based on the material residence time, the maximum historical data time is obtained through normal distribution estimation.

3. The soft sensing method of melt index based on classification and data enhancement according to claim 1, characterized in that: The classification model is an ANN model.

4. The method for soft sensing of melt index based on classification and data enhancement according to claim 1, characterized in that: The division into multiple intervals based on the production grade comprises the following steps: The sample set is divided into a plurality of intervals by constructing a pellet melt index distribution histogram of the sample set.

5. The method for soft sensing of melt index based on classification and data enhancement according to claim 1, characterized in that: In the melt index soft measurement process, after obtaining the predicted melt index of pellets at the discharge port of the extrusion device, the process further includes: Based on the obtained laboratory-measured pellet melt index and the predicted pellet melt index at the outlet of the extrusion device, a corrected pellet melt index is obtained through correction.

6. The method for soft sensing of melt index based on classification and data enhancement according to claim 1, characterized in that: The sensor data include the extrusion equipment motor torque and speed, gear pump torque and speed, powder feed rate, peroxide feed rate, extruder barrel temperature, extruder barrel pressure and throttle valve opening.

7. An electronic device, characterized in that: include: One or more processors and a memory, wherein the memory stores one or more programs, wherein the one or more programs include instructions for executing the soft-sensing method for melt index based on classification and data enhancement as claimed in any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that: The invention comprises one or more programs for execution by one or more processors of an electronic device, wherein the one or more programs comprise instructions for executing the soft-sensing method for melt index based on classification and data enhancement as claimed in any one of claims 1 to 6.

Images