Polymer multi-component content prediction method based on multi-branch neural network

Abstract

The invention discloses a polymer multi-component content prediction method based on a multi-branch neural network, and the method comprises the steps: collecting blending data of a plurality of polymers in real time in the processing process of a twin-screw extruder, carrying out the time sequence alignment of the collected different data sources to form a sample, then carrying out the preprocessing and normalization, and carrying out the prediction of the multi-component content of the polymer. Corresponding labeling information is added; a multi-branch neural network model is constructed, in the training stage, features of different data sources are mapped into a high-dimensional feature space through a branch feature extraction layer, corresponding data feature distribution is extracted, information from multiple data sources is integrated through a feature fusion module, and various feature representations are constructed; and inputting test data into the trained model for testing, calculating the difference between the component content value measured by the model and the actually set component content value, and verifying the measurement precision and stability of the model through contrastive analysis, thereby comprehensively evaluating the effectiveness and applicability of the model.

Description

Technical Field

[0001] The present invention relates to the technical field of on-line monitoring in the polymer extrusion process, and particularly relates to a method for predicting the multi-component content of polymers based on a multi-branch neural network. Background Art

[0002] The co-extrusion technology has become the main research object in the field of polymer material processing due to its characteristics such as continuous and batch production. It integrates multiple batch processing units, such as melting, mixing, homogenization, etc., and can perform different process steps in the continuous manufacturing process. Combining the advantages of on-line process analysis technology and continuous manufacturing, real-time monitoring of the product quality can be achieved, especially the accurate monitoring of the component content. The component content is one of the key factors affecting the final performance and product quality of the blended polymer. Its change can indirectly reflect the dispersion uniformity of the product, and thus can be used as an important indicator for evaluating the product quality. Therefore, accurately and quickly measuring the change of the polymer component content during the processing has important and far-reaching significance for improving the uniformity of the product quality and production efficiency, and further ensuring the high-quality and efficient application of the polymer materials.

[0003] A convolutional neural network generally refers to a computational model composed of multiple neurons connected to each other. Its basic idea originates from the simulation of the working mechanism of human brain neurons. By constructing a hierarchical network structure composed of an input layer, a hidden layer, and an output layer, the neural network can process external input data, capture the non-linear relationships and complex patterns between the data. By iteratively optimizing the connection weights between the neurons, the neural network can automatically learn the key features and potential laws from the data, so as to achieve various tasks such as classification, prediction, and optimization of the input data. Therefore, the multi-component content measurement method based on a multi-branch neural network does not require complex mechanism analysis. By automatically extracting the effective features of the data by the neural network for component content measurement, it promotes the development and application of on-line measurement of extruders.

[0004] With the rapid development of computer technology and data acquisition technology, a large amount of production process data can be obtained at the extrusion processing site, laying an important foundation for the online monitoring and intelligent development of extruders. However, there are usually multiple data sources in the extrusion process, which may include raw material characteristic data, process parameter data, equipment operation status data, and environmental variable data, etc. These various types of data not only have significant differences in the acquisition methods and dimensions, but also have complex characteristics such as dynamic changes, non-linear correlations, and multi-scale distributions. In addition, since extrusion processing involves multiple material systems, the characteristics of different materials and the changes in their mixing ratios are complex, which further increases the difficulty of constructing a model with strong generality and wide adaptability. Therefore, how to effectively integrate multi-source data, give full play to their complementary advantages, and enable them to adapt to a wider range of material systems has become the key direction for realizing intelligent and precise monitoring of the extrusion process.

[0005] An existing method for the combined real-time online monitoring and source analysis of atmospheric heavy metal particulate matter (CN201910349577.6) realizes the emission monitoring of the mass concentration of various heavy metal particles in the atmosphere and the analysis of the sources of target trace substances at the measurement points through an aerosol mass spectrometer and an X-ray fluorescence spectrometer. However, this method only realizes the combination of two types of spectrometers, does not fully combine the rich characteristics of multi-source data, and its application scope is also limited to the real-time monitoring of heavy metal particles in the atmosphere. Summary of the Invention

[0006] To solve the above technical problems, the present invention proposes a method for predicting the multi-component content of polymers based on a multi-branch neural network. The method of the present invention constructs a multi-branch network architecture, maps multi-dimensional information from different data sources to their respective high-dimensional feature spaces, thereby extracting and characterizing the feature distributions of the data, and then uses a feature fusion module to effectively integrate the multi-source data, extracts key information to the greatest extent, utilizes the information complementarity between different data sources, significantly enhances the comprehensiveness and feature expression ability of the processing process data. Finally, through a fully connected layer, the fused features are modeled and regression analyzed to accurately output the component content values of different polymers. The method of the present invention can not only improve the measurement accuracy, but also enhance the applicability of the system in different material systems, providing reliable technical support for the component content monitoring of the polymer processing process.

[0007] The present invention is realized by at least one of the following technical solutions.

[0008] A method for predicting the multi-component content of polymers based on a multi-branch neural network, comprising the following steps:

[0009] (1), Use sensors on the die head of a twin-screw extruder to collect data from multiple data sources;

[0010] (2) Preprocess the collected data;

[0011] (3) Input the preprocessed data into the trained multi-branch neural network model to predict the component contents of different material systems.

[0012] Furthermore, the data collected in step (1) includes spectral data, and the spectral data includes near-infrared spectra and Raman spectral data.

[0013] Furthermore, the preprocessing includes removing redundant parts of the data, noise reduction, baseline correction, and normalization.

[0014] Furthermore, the construction and training of the multi-branch neural network model include the following steps:

[0015] Step 1: Use sensors on the die head of a twin-screw extruder to collect data in real time, integrate the data collected by different sensors to generate samples, and label the samples, so as to obtain a sample dataset {x, y}, where x is the collected sample data and y is the corresponding component content label;

[0016] Step 2: Preprocess the sample dataset {x, y} to generate a preprocessed dataset {x p , y p}}, where x p is the preprocessed sample data and y p is the corresponding component content label; divide the preprocessed data into a training set, a validation set, and a test set;

[0017] Step 3: Construct a multi-branch neural network model for identifying multi-component contents, including: an input layer, a branch feature extraction layer, a feature fusion layer, and an output layer. Among them, the input layer is used to receive the preprocessed near-infrared and Raman spectral data; the branch feature extraction layer performs high-dimensional feature extraction on different data sources respectively, and one data source corresponds to one branch feature extraction layer; the feature fusion layer effectively integrates the feature information extracted by each branch; the output layer is used to generate the measured values of the component contents in the polymer;

[0018] Step 4: Use the training set, adopt a supervised learning method, and combine the Adam optimization algorithm to adaptively adjust the learning rate to optimize and train the parameters of each branch feature extraction layer and the output layer;

[0019] Step 5: In the test stage, input the test samples into the trained multi-branch neural network model to generate corresponding component content measured values, and compare the generated component content measured values with the set true component contents to evaluate the measurement performance of the multi-branch neural network model.

[0020] Further, in step 1, the acquisition time includes near-infrared spectroscopy and Raman spectroscopy data. Taking the Raman spectroscopy with the longest acquisition time as the time reference, the near-infrared spectroscopy is corrected for time series through data alignment technology.

[0021] Further, in step 2, the preprocessing is specifically as follows: for the near-infrared spectroscopy, it is first denoised by the Savitzky-Golay smoothing algorithm, then baseline correction and scattering correction are performed, and finally normalization processing is carried out; for the Raman spectroscopy, the data is first trimmed to remove redundant parts, then the Savitzky-Golay smoothing algorithm is applied for denoising, and then baseline correction and normalization processing are performed.

[0022] Further, each branch feature extraction layer includes multiple feature extraction units, and the multiple feature extraction units are arranged in series in sequence. Each feature extraction unit includes a batch normalization layer, a convolutional layer with a one-dimensional convolutional kernel, and a max pooling layer with a one-dimensional pooling kernel.

[0023] Further, the branch feature extraction layers respectively extract features from different data sources, and the extracted high-dimensional features are integrated for information in the feature fusion layer. The fully connected output layer uses the ReLU activation function to generate the result.

[0024] Further, the loss function of the multi-branch neural network model is the mean square error MSE loss function, which is used to measure the difference between the measured value and the true value of the multi-branch neural network model:

[0025]

[0026] In the formula, N represents the total number of samples, y i represents the true value of the i-th sample, represents the predicted value of the i-th sample;

[0027] Meanwhile, the coefficient of determination R 2 is used as a performance evaluation index to measure the data interpretation ability of the multi-branch neural network model:

[0028]

[0029] In the formula, represents the predicted result of the i-th sample, y i represents the true value of the i-th sample, represents the mean value of the true values, and n represents the total number of samples.

[0030] A computer device of the present invention includes: a memory, a processor, and a computer program stored on the memory. When the computer program is executed on the processor, the polymer multi-component content prediction method based on the multi-branch neural network as described above is implemented.

[0031] Compared with the prior art, the beneficial effects of the present invention are at least as follows:

[0032] 1. The present invention constructs a feature extraction layer using one-dimensional convolutional kernels and one-dimensional pooling kernels, avoiding the dependence on artificial feature extraction in traditional methods, significantly reducing the requirements for signal processing and related professional knowledge. Through the automated feature extraction mechanism, it can efficiently extract high-dimensional features from raw data, providing a more accurate and rich feature representation for subsequent modeling and analysis.

[0033] 2. The present invention designs a multi-branch neural network structure to evaluate and extract the feature information of materials from multiple angles, enabling it to adapt to various material systems. In some cases, when the signal response of a specific material to a certain sensor is weak, information can be supplemented through the data of other sensors, giving full play to the complementary advantages of multiple data sources. This method can comprehensively utilize the data sources of multiple sensors, significantly improving the generalization ability and application scope of the model.

[0034] 3. The network model constructed by the present invention fully considers the actual needs of industrial applications, is not limited to the feature information of a single data source, but supports the fusion of multiple data sources. In addition to spectral feature information, other types of data sources (such as process parameters like temperature and pressure) can also be combined to achieve more types of tasks, further broadening the application scenarios and practicality of the model.

[0035] 4. The present invention effectively solves the problem of measuring the multi-component content of polymers by combining multiple data sources and deeply mining the potential information in the data. This method provides a feasible technical solution for the online monitoring and intelligentization of twin-screw extruders, laying a foundation for achieving efficient and accurate industrial production control. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 is a flowchart of the training process of the multi-branch neural network model in the embodiment;

[0037] Figure 2 is a flowchart of the testing process of the multi-branch neural network model in the embodiment;

[0038] Figure 3 is a schematic diagram of the network framework of the multi-branch neural network model of the method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0039] In order to make the technical solutions and objectives of the present invention clearer and more understandable, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific implementation steps described here are only used to better illustrate the application of the present invention, but the technical features involved in the embodiments of the present invention are not limited thereto.

[0040] Online prediction method for polymer multi-component content based on multi-branch neural network, comprising the following steps:

[0041] (1) Real-time collect spectral data or other types of data sources (such as process parameters like temperature, pressure, etc.) using sensors on the die head of a twin-screw extruder.

[0042] (2) Preprocess the collected data.

[0043] (3) Input the preprocessed data into a trained multi-branch neural network model to predict the component content of different material systems.

[0044] The construction and training of the multi-branch neural network model include the following steps:

[0045] Step 1: Collect near-infrared spectral data and Raman spectral data of the extruded products of the extruder, integrate the data collected by different sensors to generate samples, and label the samples to obtain a large number of sample datasets. The sample datasets contain different types of data and corresponding labels.

[0046] Step 2: Perform data preprocessing, preprocess the dataset constructed in Step 1 to generate a preprocessed dataset, and divide the preprocessed data into a training set, a validation set, and a test set.

[0047] Step 3: Construct a multi-branch neural network model. The multi-branch neural network model includes: an input layer, a branch feature extraction layer, a feature fusion layer, and an output layer. Among them, the input layer is used to receive the preprocessed data; the branch feature extraction layer performs high-dimensional feature extraction on different data sources respectively; the feature fusion layer effectively integrates the feature information extracted by each branch; the output layer is used to generate the measured values of the component content in the polymer.

[0048] In view of the characteristic that the input data are all one-dimensional signals, one data source corresponds to one branch feature extraction layer. Each branch feature extraction layer includes multiple feature extraction units, and the multiple feature extraction units are arranged in series in sequence. Each feature extraction unit includes a batch normalization layer, a convolutional layer with a one-dimensional convolutional kernel, and a max pooling layer with a one-dimensional pooling kernel.

[0049] After flattening the outputs of each branch feature extraction layer and then splicing them, input them into the output layer with a fully connected layer structure. Its output layer all adopts the ReLU output function, and its output labels are {1, 2, 3,..., m}, where m is the component content of different materials respectively.

[0050] Step 4: Use the training set in Step 2 to optimize and train the parameters of the branch feature extraction layer and the parameters of the fully connected layer using traditional supervised learning algorithms and the Adam algorithm.

[0051] In some embodiments of the present invention, the regression loss function of the output layer is the mean squared error loss function (MSE):

[0052]

[0053] where N represents the total number of samples, y i represents the true value of the i-th sample, represents the predicted value of the i-th sample.

[0054] The coefficient of determination R 2 is used as a performance evaluation index to measure the model's ability to explain the data:

[0055]

[0056] where represents the prediction result of the i-th sample, y i represents the true value of the i-th sample, represents the mean of the true values, and n represents the total number of samples.

[0057] In the test stage, by outputting the predicted values of the component contents of the test samples and comparing them with the actual values, the RMSE and R of the model are calculated 2 as evaluation indexes to evaluate the prediction performance and accuracy of the model.

[0058] Taking the acquisition of near-infrared spectroscopy and Raman spectroscopy data as an example, the online prediction method for polymer multi-component contents based on a multi-branch neural network in this embodiment includes the following steps:

[0059] (1) Taking the acquisition of near-infrared spectroscopy and Raman spectroscopy data by the sensors on the die head of a twin-screw extruder as an example.

[0060] (2) Preprocessing the acquired data.

[0061] The preprocessing in this embodiment is specifically as follows: the near-infrared spectroscopy is first denoised by the Savitzky-Golay smoothing algorithm, then baseline correction and scattering correction are performed, and finally normalization processing is carried out; the Raman spectroscopy is first subjected to data clipping to remove redundant parts, then the Savitzky-Golay smoothing algorithm is applied for denoising, and then baseline correction and normalization processing are carried out.

[0062] (3) Inputting the preprocessed data into the trained multi-branch neural network model to predict the component contents of different material systems.

[0063] As Figure 1 、 Figure 2 shown, the construction and training of the multi-branch neural network model in this embodiment include the following steps:

[0064] Step 1: Collect the near-infrared spectrum data and Raman spectrum data of the products extruded by the extruder, integrate the data collected by different sensors to generate samples, and label the samples to obtain a large number of sample datasets {x, y}. The sample datasets include near-infrared spectra, Raman spectra, and corresponding labels, where x is the sample dataset composed of the two spectral data of near-infrared spectra and Raman spectra, and y is the corresponding component content label.

[0065] During the collection process, aiming at the problems of inconsistent time series and different acquisition frequencies existing in different data sources, taking the Raman spectrum with the longest acquisition time as the time reference, the near-infrared spectrum is corrected for time series through data alignment technology to ensure the consistency of multi-data sources in the time dimension.

[0066] Step 2: Perform data preprocessing, preprocess the dataset constructed in Step 1 to generate a preprocessed dataset {x p , y p}}, where x p is the preprocessed sample data, and y p is the corresponding component content label; divide the preprocessed data into a training set, a validation set, and a test set.

[0067] The specific preprocessing is as follows: The near-infrared spectrum is first denoised by the Savitzky-Golay smoothing algorithm, then baseline correction and scattering correction are performed, and finally normalization processing is carried out. The Raman spectrum is first trimmed to remove redundant parts, then the Savitzky-Golay smoothing algorithm is applied for denoising, and then baseline correction and normalization processing are carried out.

[0068] Step 3: Construct a multi-branch neural network model. The multi-branch neural network model includes: an input layer, a branch feature extraction layer, a feature fusion layer, and an output layer. Among them, the input layer is used to receive the preprocessed near-infrared and Raman spectrum data; the branch feature extraction layer performs high-dimensional feature extraction on different data sources respectively; the feature fusion layer effectively integrates the feature information extracted by each branch; the output layer is used to generate the measured values of the component contents in the polymer.

[0069] In some embodiments of the present invention, aiming at the characteristic that the input data are all one-dimensional signals, one data source corresponds to one branch feature extraction layer, each branch feature extraction layer includes multiple feature extraction units, the multiple feature extraction units are arranged in series in sequence, and each feature extraction unit includes a batch normalization layer, a convolutional layer with a one-dimensional convolutional kernel, and a max-pooling layer with a one-dimensional pooling kernel.

[0070] Such as Figure 3As shown, in some embodiments of the present invention, after flattening the outputs of each branch feature extraction layer, the data of multiple data sources are concatenated along the feature dimension and input into an output layer with a fully connected layer structure. The output layer all adopts the ReLU output function, and its output labels are {1, 2, 3,..., m}, where m is the component content of each different material.

[0071] In this embodiment, the near-infrared spectrum and the Raman spectrum are respectively input into different branch feature extraction layers. The branch feature extraction layer performs independent feature extraction on the near-infrared spectrum and the Raman spectrum respectively. Through the hierarchical processing of multiple feature extraction units, each branch can extract deep high-dimensional features of its corresponding spectral data, so as to fully mine and characterize the feature information of each data source. Please refer to Figure 2 , in this embodiment, 3 feature extraction units are set. The size of the first convolutional kernel of the first feature extraction unit of the near-infrared branch feature extraction layer is set to 40, and the size of the maximum pooling layer is set to 2. The convolutional kernel sizes of the second and third feature extraction units are set to 3, and the size of the maximum pooling layer is set to 2. The size of the first convolutional kernel of the first feature extraction unit of the Raman branch feature extraction layer is set to 40, and the size of the maximum pooling layer is set to 2. The convolutional kernel sizes of the second and third feature extraction units are set to 3, and the size of the maximum pooling layer is set to 2. The branch feature extraction layer respectively extracts features from the near-infrared spectrum and the Raman spectrum. These extracted high-dimensional features are then effectively integrated in the feature fusion layer, and the complementary nature between multiple data sources is utilized to achieve deep fusion of information, forming a unified feature representation. The fused features are modeled through a fully connected layer, and the fully connected layer adopts the ReLU activation function to ensure that the model has the ability of non-linear expression and generate output results at the same time.

[0072] Step 4: Using the training set in Step 2, optimize and train the parameters of the branch feature extraction layer and the parameters of the fully connected layer by using traditional supervised learning algorithms and the Adam algorithm.

[0073] In some of the embodiments of the present invention, the regression loss function of the output layer is the mean squared error loss function (Mean Squared Error, MSE):

[0074]

[0075] In the formula, N represents the total number of samples, y i represents the true value of the i-th sample, and y i represents the predicted value of the i-th sample.

[0076] It should be noted that the model is first supervised and trained with a relatively large learning rate (the learning rate is set to 0.01). When the validation set loss does not change significantly for 10 epochs, the learning rate is halved (the learning rate is set to 0.005), and the minimum lower limit of the learning rate is set to 0.00001, which can accelerate the convergence of the network.

[0077] Use the coefficient of determination R 2 as a performance evaluation index to measure the ability of the model to explain the data:

[0078]

[0079] In the formula, represents the prediction result of the i-th sample, and y i represents the true value of the i-th sample, represents the mean of the true values and represents the total number of samples.

[0080] In the test stage, by outputting the predicted values of the component contents of the test samples and comparing them with the actual values, the RMSE and R of the model are calculated 2 as evaluation indexes to evaluate the prediction performance and accuracy of the model.

[0081] The present invention will be further described below in conjunction with the accompanying drawings and experimental cases.

[0082] To evaluate the performance of the proposed method, in some embodiments of the present invention, different component contents of three polymers, namely polypropylene (PP), polyolefin elastomer (POE), and polyethylene (PE), are collected respectively. The data of the three polymers are collected on a twin-screw extrusion die head. The screw speed is set to 50 rpm, the feeding speed is 1 rpm, and the collection time for each group is 20 minutes. It should be noted that to avoid interference between materials of different components, before conducting the next group of experiments, the barrel needs to be completely emptied and the equipment needs to be cleaned according to specific process parameters. In each group of experiments, the actual effective data collection time is 10 minutes. Among them, the collection period of near-infrared is 1.5 s, and the collection period of Raman is 6 s. The temperature zones are set to 120 °C, 160 °C, 180 °C, 200 °C, 200 °C, 200 °C, 210 °C, 210 °C, 210 °C, 200 °C.

[0083] To construct the training set, validation set, and test set of the multi-branch neural network, there are 60 pieces of data for each material combination, with a total of 45 combinations. The dimension of the near-infrared spectral features is 512, and the dimension of the Raman spectral features is 1024. To avoid the interference of redundant information, the Raman spectral data is intercepted, and 650-dimensional data is intercepted from the Raman spectrum. Ten of them are selected as the test set, and the remaining data is divided into the training set and the validation set according to the ratio of 6:4.

[0084] Table 1 Different component contents of PP\POE\PE

[0085] Sample number PP POE PE 1 8 0 0 2 7 1 0 3 6 2 0 4 5 3 0 5 4 4 0 6 0 5 0 7 2 6 0 8 1 7 0 9 0 8 0 10 0 7 1 11 0 6 2 12 0 5 3 13 0 4 4 14 0 3 5 15 0 2 6 16 0 1 7 17 0 0 8 18 1 0 7 19 2 0 6 20 3 0 5 21 4 0 4 22 5 0 3 23 6 0 2 24 7 0 1 25 6 1 1 26 5 1 2 27 4 1 3 28 3 1 4 29 2 1 5 30 1 1 6 31 1 2 5 32 1 3 4 33 1 4 3 34 1 5 2 35 1 6 1 36 2 5 1 37 3 4 1 38 4 3 1 39 5 2 1 40 4 2 2 41 3 3 2 42 2 4 2 43 2 3 3 44 3 2 3 45 2 2 4

[0086] Since the convolutional kernel has strong feature extraction ability, one-dimensional convolutional kernels are used to construct the network for feature extraction and output regression of the branches. Taking the near-infrared branch as an example, its module parameters are shown in Table 2.

[0087] Table 2 Structure of the near-infrared branch feature extraction layer

[0088]

[0089] Table 2 describes the structure of a feature extraction layer, including multiple convolutional neural network modules and their parameter configurations. Among them, the activation function represents a non-linear function used to introduce non-linear characteristics. The number represents the number of convolutional kernels, the kernel size represents the size of the convolutional kernel, the stride represents the distance that the convolutional kernel slides each time, and the output size represents the size of the tensor output by each layer and the batch size.

[0090] The input layer receives 512-dimensional data. After the first layer of convolution (Conv_1), this convolutional layer has 16 convolutional kernels, the kernel size is 40, the stride is 1, the activation function is ReLU, and the output size is (64, 512, 16). Subsequently, through the first layer of max-pooling layer (MaxPool_1), the output size is halved to (64, 256, 16), and batch normalization layer (BN_1) is performed. The second layer of convolution (Conv_2) uses 32 convolutional kernels, and the output size remains (64, 256, 32). After batch normalization (BN_2) and max-pooling (MaxPool_2), the output size becomes (64, 128, 32). The last layer of convolution (Conv_3) uses 32 convolutional kernels. After batch normalization (BN_3) and max-pooling (MaxPool_3), the final output size is (64, 64, 32).

[0091] To verify the superiority of the method of the present invention, several classical algorithms are used for comparison on the 45 groups of data in Table 1. The algorithms include convolutional neural network (CNN) and multi-layer perceptron (MLP). To avoid the contingency of the results, it is selected to take the RMSE and R after training five times. 2average value.

[0092] Table 3 Comparison of MBNN of the present invention with CNN and MLP

[0093]

[0094] According to the experimental results in Table 3, it can be seen that the multi-branch neural network model (MBNN) of the present invention is significantly better than CNN and MLP in terms of performance indicators. 2 In terms of indicators, MBNN reached 0.9083, which is significantly higher than CNN's 0.8504 and MLP's 0.6812, indicating that MBNN has a stronger ability to fit the target variable. In terms of RMSE indicators, the error of MBNN is 6.04, which is significantly lower than CNN's 7.72 and MLP's 11.27, reflecting its superiority in prediction accuracy. These results show that the multi-branch neural network model of the present invention performs well in balancing fitting ability and prediction accuracy, and provides an effective solution for the target task.

[0095] The present invention conducts in-depth research on the key problem that it is difficult to accurately identify and distinguish the content of different material components due to the similar molecular structures of the materials and the coupling relationship between the data features during the extrusion process of different polymers PP, POE, and PE by a twin-screw extruder. Taking PP, POE and PE as the research objects, the present invention designs and utilizes a multi-branch convolutional neural network structure. By processing the different feature branches of the input data and combining the fusion mechanism of the deep-level features in the network structure, this method can efficiently extract key features, thereby accurately measuring the component content in different material systems. Experiments show that this method can significantly improve the ability to measure the content of multiple components in complex industrial environments, provide technical support for real-time monitoring and precise control of the polymer extrusion process, and has good industrial application prospects.

[0096] The present invention constructs a multi-branch neural network model, which makes full use of the rich features of multiple data sources. By fusing and deeply mining heterogeneous data from different sources, it can not only accurately measure the multi-component content of polymers and achieve efficient and reliable component analysis, but also be applicable to complex scenarios such as polymer fault monitoring. This method has significant flexibility and scalability in terms of application scope. It is not only limited to data analysis in a single field, but can also be widely used in industrial quality control, material performance evaluation, and equipment operation status diagnosis. It provides a new technical solution for intelligent monitoring driven by multi-source data.

[0097] It should be noted that although the implementation of the present invention has been elaborated in detail with reference to the examples, those skilled in the art can easily understand that any modifications, substitutions, improvements, etc. made without departing from the spirit and principles of the present invention set forth in the appended claims should be included within the protection scope of the present invention.

Claims

1. A method for predicting the content of multiple components of a polymer based on a multi-branch neural network, characterized in that: The following steps are involved: (1) Using sensors on the die head of a twin-screw extruder to collect data from multiple data sources; (2) Preprocessing the collected data; (3) Input the preprocessed data into the trained multi-branch neural network model to predict the component content of different material systems.

2. The method for predicting polymer multi-component content based on a multi-branch neural network according to claim 1, characterized in that: The data collected in step (1) includes spectral data, and the spectral data includes near-infrared spectral data and Raman spectral data.

3. The method for predicting polymer multi-component content based on a multi-branch neural network according to claim 1, characterized in that: Preprocessing includes removing redundant parts of the data, noise reduction, baseline correction, and normalization.

4. The method for predicting polymer multi-component content based on a multi-branch neural network according to claim 1, characterized in that: The multi-branch neural network model construction and training includes the following steps: Step 1: Use the sensor on the die head of the twin-screw extruder to collect data in real time, integrate the data collected by different sensors to generate samples, and annotate the samples to obtain a sample data set {x, y}, where x is the collected sample data and y is the corresponding component content label; Step 2: Preprocess the sample data set {x, y} to generate the preprocessed data set {x p ,y p }, where x p is the sample data after preprocessing, y p is the corresponding component content label; the preprocessed data is divided into a training set, a validation set, and a test set; Step 3, construct a multi-branch neural network model for identifying multi-component content, including: an input layer, a branch feature extraction layer, a feature fusion layer, and an output layer, wherein the input layer is used to receive preprocessed near-infrared and Raman spectral data; the branch feature extraction layer performs high-dimensional feature extraction on different data sources respectively, and one data source corresponds to one branch feature extraction layer; the feature fusion layer effectively integrates the feature information extracted by each branch; the output layer is used to generate the measured value of the content of each component in the polymer; Step 4: Using the training set, adopting the supervised learning method, and combining the Adam optimization algorithm to adaptively adjust the learning rate, the parameters of the feature extraction layer and output layer of each branch are optimized and trained; Step 5. During the testing phase, the test sample is input into the trained multi-branch neural network model to generate corresponding component content measurement values, and the generated component content measurement values ​​are compared with the set true component content to evaluate the measurement performance of the multi-branch neural network model.

5. The method for predicting polymer multi-component content based on a multi-branch neural network according to claim 4, characterized in that: In step 1, the acquisition time includes near-infrared spectrum and Raman spectrum data. The Raman spectrum with the longest acquisition time is used as the time reference, and the near-infrared spectrum is time-corrected by data alignment technology.

6. The method for predicting polymer multi-component content based on a multi-branch neural network according to claim 5, characterized in that: In step 2, the preprocessing is as follows: the near-infrared spectrum is first denoised by the Savitzky-Golay smoothing algorithm, then baseline correction and scattering correction are performed, and finally normalization is performed; the Raman spectrum is first data trimmed to remove redundant parts, then the Savitzky-Golay smoothing algorithm is used for denoising, and then baseline correction and normalization are performed.

7. The method for predicting polymer multi-component content based on a multi-branch neural network according to claim 4, characterized in that: Each branch feature extraction layer includes multiple feature extraction units, and the multiple feature extraction units are arranged in series in sequence. Each feature extraction unit includes a batch normalization layer, a convolution layer of a one-dimensional convolution kernel, and a maximum pooling layer of a one-dimensional pooling kernel.

8. The method for predicting polymer multi-component content based on a multi-branch neural network according to claim 7, characterized in that: The branch feature extraction layer extracts features from different data sources respectively. The extracted high-dimensional features are integrated in the feature fusion layer, and the fully connected output layer uses the ReLU activation function to generate the result.

9. The method for predicting polymer multi-component content based on a multi-branch neural network according to claim 1, characterized in that: The loss function of the multi-branch neural network model is the mean square error MSE loss function, which is used to measure the difference between the measured value and the true value of the multi-branch neural network model: In the formula, N represents the total number of samples, y i represents the true value of the i-th sample, Represents the predicted value of the i-th sample; At the same time, using the coefficient of determination R 2 As a performance evaluation indicator, it measures the ability of the multi-branch neural network model to interpret data: In the formula, Represents the prediction result of the i-th sample, y i represents the true value of the i-th sample, represents the mean of the true value, and n represents the total number of samples.

10. A computer device, characterized in that: It comprises: a memory and a processor and a computer program stored in the memory. When the computer program is executed on the processor, it realizes the polymer multi-component content prediction method based on a multi-branch neural network as described in any one of claims 1 to 9.

Images