Method and device for determining rationality of feature extraction of refining process

By constructing a full-process matrix and training a multi-scale convolutional neural network model, combining a hybrid Gaussian model and a t-distribution-random neighbor embedding manifold learning, the problem of inaccurate feature extraction of the refining process flow is solved, and the refining process flow is realized Effective extraction and evaluation of features.

CN115049005BActive Publication Date: 2024-12-03CNOOC HUIZHOU PETROCHEM CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210689093.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-06-16
Publication Date
2024-12-03
Estimated Expiration
2042-06-16

Smart Images

  • Figure CN115049005B_ABST
    Figure CN115049005B_ABST
Patent Text Reader

Abstract

The present application discloses a method and device for determining the rationality of feature extraction of a refining process. The method includes: obtaining industrial historical sample data; constructing a full process matrix based on the industrial historical sample data to obtain a training set and a test set; training a multi-scale convolutional neural network model based on the training set to obtain a trained multi-scale convolutional neural network model; inputting the test set into the trained multi-scale convolutional neural network model to obtain an intermediate feature map; and determining the rationality of feature extraction of the multi-scale convolutional neural network model based on the test set and the intermediate feature map through a mixed Gaussian model and t-distribution-random neighbor embedding manifold learning. The present application is conducive to solving the problem in the prior art that it is impossible to realize feature extraction of the process flow through the fusion of multi-scale convolutional networks and process flow knowledge.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a method and device for determining the rationality of feature extraction of a refining process. Background Art

[0002] In recent years, data-driven process modeling has been increasingly widely studied and applied. Data-driven modeling does not require in-depth study of the complex mechanisms of industrial processes. By learning historical data through machine learning algorithms, fault diagnosis, parameter warning and production optimization of the production process can be achieved. However, oil refining processes such as hydrocracking involve multiple process units such as feed, reaction, separation, and fractionation. The physical and chemical mechanisms are complex, and the process data is high-dimensional, has large time lags, and is nonlinear, which brings challenges to the effective implementation of feature engineering and then machine learning modeling.

[0003] In 2006, Hinton et al. proposed the concept of deep learning and constructed deep belief networks (DBNs). They used layer-by-layer greedy unsupervised pre-training of multiple stacked restricted Boltzmann machines to extract feature representations layer by layer, and combined with supervised fine-tuning to finally predict the target variable. Another deep learning framework, auto-encoder (AE), follows the same unsupervised pre-training-supervised fine-tuning parameter-finding paradigm as DBN, and uses gradient descent (GD) to solve network parameters. It has been more widely studied and applied in recent years. Yuan et al. conducted research on process industry modeling based on AE-type deep learning algorithms and achieved a series of results. It is worth noting that as unsupervised deep learning algorithms, the features extracted by DBN, DAE, etc. include both effective features that are beneficial to target prediction and invalid features or even noise. This requires additional regularization in the pre-training stage to strengthen the identification and extraction of effective features, or supervised training in the fine-tuning stage to make all extracted features fit the optimal representation. The great success of the AlexNET model in the ImageNET image recognition competition in 2012 marked the entry of supervised deep learning represented by convolutional neural networks (CNN) into the industrial application stage. As the mainstream technology of deep learning in computer vision, CNN uses convolution kernels to learn local spatial receptive fields, thereby efficiently extracting the feature relationship between local pixels of the image, while realizing sparse connections between different layers and weight sharing between the same layers. On the other hand, CNN achieves dimensionality reduction sampling of intermediate layer features through pooling, which improves the robustness of feature extraction, allowing CNN to achieve end-to-end extraction of highly robust features while maintaining a low amount of parameter calculations. The various advantages of CNN have made it an important algorithm for current deep learning in process industry modeling. Wang et al. conducted a series of studies on the shape and convolution method of CNN convolution kernels and verified them on methanol synthesis industrial equipment. Geng et al. used cross feature engineering (CF) to improve the feature dimension of data samples, and proposed a cross feature convolution network (CF-CNN) that was successfully used for energy efficiency and carbon emission analysis optimization in the process of naphtha cracking to produce ethylene.

[0004] Most of the above CNN process modeling research starts from sample data, and obtains sample data of different scale levels by adjusting the sample data structure or sampling preprocessing to achieve multi-scale feature extraction. However, there are still deficiencies in CNN's multi-scale convolution research on process flow and process flow knowledge fusion. Therefore, the existing technology has the problem of being unable to achieve feature extraction of process flow through multi-scale convolutional networks and process flow knowledge fusion. Summary of the invention

[0005] The purpose of the embodiments of the present application is to provide a method and device for determining the rationality of feature extraction of a refining process flow, so as to solve the problem in the prior art that feature extraction of the process flow cannot be achieved through multi-scale convolutional networks and process flow knowledge fusion.

[0006] In order to achieve the above-mentioned purpose, the first aspect of the present application provides a method for determining the rationality of feature extraction of a refining process, the method comprising:

[0007] Obtain industrial historical sample data;

[0008] Construct a full-process matrix based on industrial historical sample data to obtain training sets and test sets;

[0009] Training a multi-scale convolutional neural network model according to the training set to obtain a trained multi-scale convolutional neural network model;

[0010] Input the test set into the trained multi-scale convolutional neural network model to obtain the intermediate feature map;

[0011] Through mixed Gaussian model and t-distribution-random neighbor embedding manifold learning, the rationality of feature extraction of multi-scale convolutional neural network model is determined based on test sets and intermediate feature maps.

[0012] In the embodiment of the present application, a full process matrix is ​​constructed based on industrial historical sample data to obtain a training set and a test set including:

[0013] The initial full-process matrix is ​​obtained based on industrial historical data;

[0014] Randomly rearrange and shuffle the initial full-process matrix to obtain a shuffled full-process matrix;

[0015] The shuffled full-process matrix is ​​processed to obtain the training set and the test set.

[0016] In the embodiment of the present application, the shuffled full-process matrix is ​​processed to obtain a training set and a test set including:

[0017] Performing outlier elimination processing on the data of the shuffled full-process matrix to obtain outlier-eliminated data;

[0018] Normalize the data after outlier removal to obtain sample data;

[0019] Processing the sample data according to the combined feature construction method to obtain processed sample data;

[0020] The processed sample data is divided into training set and test set according to the data set segmentation method.

[0021] In the embodiment of the present application, outlier removal processing is performed on the data of the shuffled full-process matrix to obtain data after outlier removal processing, including:

[0022] The historical data involving shutdowns and abnormal production conditions are eliminated from the shuffled full-process matrix data.

[0023] In the embodiment of the present application, the data after outlier removal is normalized to satisfy formula (1):

[0024]

[0025] Among them, j represents the j-th dimension feature column, χ j is the sample data of the j-th dimension feature, is the sample mean of the j-th dimension feature, χ j,max is the maximum value of the j-th dimension feature, χ j,min is the minimum value of the j-th dimension feature.

[0026] In the embodiment of the present application, the sample data is processed according to the combined feature construction method to obtain the processed sample data including:

[0027] Obtain original sample features of different dimensions of sample data;

[0028] Process calculations are performed on the original sample features of different dimensions to obtain the final features;

[0029] The processed sample data is obtained according to the final features.

[0030] In an embodiment of the present application, dividing the processed sample data into a training set and a test set according to a data set segmentation method includes:

[0031] Perform working condition analysis on sample data to obtain different working condition data sets;

[0032] The working condition data set is divided into a training set and a test set.

[0033] In an embodiment of the present application, the same layer of the neural network in the multi-scale convolutional neural network model performs dilated convolution through convolution kernels of different sizes and dilated convolution kernels of different dilation rates.

[0034] In the embodiment of the present application, by using a mixed Gaussian model and t-distribution-random neighbor embedding manifold learning, the rationality of feature extraction of a multi-scale convolutional neural network model is determined based on a test set and an intermediate feature map, including:

[0035] Cluster the predictor variables in the test set according to the Gaussian mixture model to obtain the corresponding Gaussian mixture model labels;

[0036] The mixed Gaussian model labels are concatenated with the test set and the intermediate feature map to obtain the data to be reduced in dimension;

[0037] Perform dimensionality reduction visualization on the dimensionality-reduced data according to t-distribution-random neighbor embedding to obtain two-dimensional feature clusters;

[0038] The rationality of feature extraction of multi-scale convolutional neural network models is determined based on two-dimensional feature clusters.

[0039] The second aspect of the present application provides a device for determining the rationality of feature extraction of a refining process, characterized in that it includes:

[0040] a memory configured to store instructions; and

[0041] A processor is configured to call instructions from a memory and implement any of the above methods for determining the rationality of refining process flow feature extraction when executing the instructions.

[0042] Through the above technical scheme, first obtain industrial historical sample data, then build a full process matrix based on the industrial historical sample data to obtain a training set and a test set; train a multi-scale convolutional neural network model based on the training set to obtain a trained multi-scale convolutional neural network model; further input the test set into the trained multi-scale convolutional neural network model to obtain an intermediate feature map; finally, through a mixed Gaussian model and t-distribution-random neighbor embedding manifold learning, according to the test set and the intermediate feature map, determine the rationality of the feature extraction of the multi-scale convolutional neural network model. This application extracts features from the refining process flow through a multi-scale convolutional neural network, and evaluates the rationality of the features extracted by the multi-scale convolutional neural network based on a mixed Gaussian model and t-distribution-random neighbor embedding popular learning, thereby verifying the effectiveness of the multi-scale convolutional neural network in extracting features from the refining process flow, which is beneficial to solving the problem in the prior art that the feature extraction of the process flow cannot be achieved through the fusion of multi-scale convolutional networks and process flow knowledge.

[0043] Other features and advantages of the embodiments of the present application will be described in detail in the subsequent specific implementation section. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the following specific implementations, they are used to explain the embodiments of the present application, but do not constitute a limitation on the embodiments of the present application. In the accompanying drawings:

[0045] Figure 1 A schematic flow chart of a method for determining the rationality of feature extraction of a refining process flow according to an embodiment of the present application;

[0046] Figure 2 Schematic diagram of a GMM-t-SNE feature evaluation framework according to an embodiment of the present application;

[0047] Figure 3 A schematic flow chart of a method for determining the rationality of feature extraction of a hydrocracking process according to an embodiment of the present application;

[0048] Figure 4 It is a schematic diagram of a matrix design based on hydrocracking process flow knowledge according to a specific embodiment of the present application;

[0049] Figure 5 is a schematic diagram of data enhancement according to a specific embodiment of the present application;

[0050] Figure 6(a) to Figure 6(d) is a schematic diagram of GMM cluster analysis according to a specific embodiment of the present application;

[0051] Figure 7(a) to Figure 7(h) A schematic diagram of the prediction of hydrocracking process variables according to a specific embodiment of the present application;

[0052] Figure 8(a) to Figure 8(j) is a schematic diagram of t-SNE visualization of IFM according to a specific embodiment of the present application;

[0053] Fig. 9 The present invention is a structural block diagram of a device for determining the rationality of feature extraction of a refining process according to an embodiment of the present application. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. It should be understood that the specific implementation methods described herein are only used to illustrate and explain the embodiments of the present application, and are not used to limit the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0055] It should be noted that if the embodiments of the present application involve directional indications (such as up, down, left, right, front, back...), the directional indications are only used to explain the relative position relationship, movement status, etc. between the components under a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indication will also change accordingly.

[0056] In addition, if there are descriptions involving "first", "second", etc. in the embodiments of the present application, the descriptions of "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or suggesting their relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In addition, the technical solutions between the various embodiments can be combined with each other, but they must be based on the ability of ordinary technicians in the field to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such combination of technical solutions does not exist and is not within the scope of protection required by this application.

[0057] Figure 1 FIG. 1 is a flow chart of a method for determining the rationality of feature extraction of a refining process according to an embodiment of the present application. Figure 1 As shown, the embodiment of the present application provides a method for determining the rationality of refining process feature extraction, which may include the following steps:

[0058] Step 101: Obtain industrial historical sample data;

[0059] Step 102: construct a full-process matrix based on the industrial historical sample data to obtain a training set and a test set;

[0060] Step 103: training a multi-scale convolutional neural network model according to the training set to obtain a trained multi-scale convolutional neural network model;

[0061] Step 104: input the test set into the trained multi-scale convolutional neural network model to obtain an intermediate feature map;

[0062] Step 105: Determine the rationality of feature extraction of the multi-scale convolutional neural network model based on the test set and the intermediate feature map through mixed Gaussian model and t-distribution-random neighbor embedding manifold learning.

[0063] In an embodiment of the present application, industrial historical sample data of the refining process is first obtained. The complexity of different refining process flows is not the same. The entire refining process flow can be composed of multiple local sub-processes, and each local sub-process may be self-built and there may be a process coupling relationship. Therefore, when obtaining industrial historical sample data, the data of the operating units of each local sub-process of the refining process flow should be covered. A full-process matrix is ​​constructed based on the acquired industrial historical sample data, and the full-process matrix is ​​divided into a training set and a test set. First, an initial full-process matrix is ​​constructed based on the industrial historical sample data acquired by the processor, and then the sample data in the constructed initial full-process matrix is ​​shuffled, and the shuffled full-process matrix is ​​sequentially processed for outlier removal, data normalization, and combined feature construction, and the processed full-process matrix is ​​further segmented into data sets to obtain training sets and test sets.

[0064] Next, a supervised training task is constructed based on the training set, and the multi-scale convolutional neural network model is trained to obtain a trained multi-scale convolutional neural network model. Among them, the multi-scale convolutional neural network model is proposed based on the multi-scale convolution of the Inception structure block, and the process matrix designed according to the characteristics of the refining process is used as the carrier of knowledge fusion, and then a multi-scale convolutional network is established based on the novel extended-Inception convolutional structure block to complete the construction of the MSCNN algorithm. Considering that the multi-scale convolution based on the Inception structure block can significantly improve the feature extraction capability of a single convolutional layer and help to extract the feature relationship between various variables between local process flows, this application extends the Inception structure block (Extended-Inception) and introduces dilated convolution into the multi-scale convolutional neural network, which is beneficial to improving the feature extraction capability of the multi-scale convolutional neural network.

[0065] Next, the test set is input into the trained multi-scale convolutional neural network model to obtain the intermediate feature map. In one example, the target variable of the test set is predicted by the trained multi-scale convolutional neural network, and the intermediate feature map (IFM) is output to fully extract the characteristic relationship of the local-global process flow, thereby achieving accurate prediction of the predicted variable. Finally, the rationality of the feature extraction of the multi-scale convolutional neural network model is determined based on the test set and the intermediate feature map through the mixed Gaussian model and t-distribution-random neighbor embedding manifold learning. In one example, the target variables (labels) of the test set can be clustered by the mixed Gaussian model (GMM) to obtain the corresponding GMM labels; then the GMM labels are spliced ​​with the test set and IFM respectively to form the data to be reduced in dimension; further, the data to be reduced in dimension is visualized based on the t-distribution-random neighbor embedding (t-SNE), and the rationality of the feature extraction of the multi-scale convolutional neural network is evaluated on this basis.

[0066] Through the above technical scheme, first obtain industrial historical sample data, then build a full process matrix based on the industrial historical sample data to obtain a training set and a test set; train a multi-scale convolutional neural network model based on the training set to obtain a trained multi-scale convolutional neural network model; further input the test set into the trained multi-scale convolutional neural network model to obtain an intermediate feature map; finally, through a mixed Gaussian model and t-distribution-random neighbor embedding manifold learning, according to the test set and the intermediate feature map, determine the rationality of the feature extraction of the multi-scale convolutional neural network model. This application extracts features from the refining process flow through a multi-scale convolutional neural network, and evaluates the rationality of the features extracted by the multi-scale convolutional neural network based on a mixed Gaussian model and t-distribution-random neighbor embedding popular learning, thereby verifying the effectiveness of the multi-scale convolutional neural network in extracting features from the refining process flow, which is beneficial to solving the problem in the prior art that the feature extraction of the process flow cannot be achieved through the fusion of multi-scale convolutional networks and process flow knowledge.

[0067] In the embodiment of the present application, step 102, constructing a full process matrix according to the industrial historical sample data to obtain a training set and a test set may include:

[0068] The initial full-process matrix is ​​obtained based on industrial historical sample data;

[0069] Randomly rearrange and shuffle the initial full-process matrix to obtain a shuffled full-process matrix;

[0070] The shuffled full-process matrix is ​​processed to obtain the training set and the test set.

[0071] Specifically, the processor constructs an initial full-process matrix based on the acquired industrial historical sample data. In order to improve the richness of the data arrangement in the initial full-process matrix, the initial full-process matrix can be randomly rearranged and shuffled to obtain a shuffled full-process matrix, thereby expanding the sample data of the full-process matrix and achieving data enhancement. Further, the shuffled full-process matrix is ​​processed, and the shuffled full-process matrix is ​​sequentially subjected to outlier removal, data normalization, and combined feature construction processing to finally obtain the processed full-process matrix, and the sample data in the processed full-process matrix is ​​divided into a training set and a test set according to the data set segmentation method.

[0072] Taking the hydrocracking process as an example, the hydrocracking process is highly complex, and there are process coupling relationships within and between local processes. The local sub-processes of the hydrocracking process are taken as rows, and the process variables contained in each sub-process are taken as columns. A total of 126 key process variables in the industrial historical sample data are selected to construct a 9×14 initial full-process matrix. Then, the initial full-process matrix is ​​subjected to random rearrangement techniques in the row direction. The richness of the sub-process arrangement is improved through "row shuffling", and the purpose of sample enhancement is achieved at the same time. Then, the shuffled full-process matrix is ​​processed. Finally, according to the method of data set segmentation, the sample data in the processed full-process matrix is ​​divided into a training set and a test set at a ratio of 80%:20%. The random row rearrangement and shuffling process is conducive to improving the richness of sample data and achieving the purpose of data enhancement.

[0073] In an embodiment of the present application, processing the shuffled full-process matrix to obtain a training set and a test set may include:

[0074] Performing outlier elimination processing on the data of the shuffled full-process matrix to obtain outlier-eliminated data;

[0075] Normalize the data after outlier removal to obtain sample data;

[0076] Processing the sample data according to the combined feature construction method to obtain processed sample data;

[0077] The processed sample data is divided into training set and test set according to the data set segmentation method.

[0078] Specifically, after shuffling, the sample data in the full-process matrix is ​​multiplied, and while the data strength is enhanced, the data complexity is also increased. Therefore, the shuffled full-process matrix needs to be processed to further obtain the training set and the test set. First, the data involving shutdowns and abnormal production conditions in the full-process matrix are filtered out by outlier elimination, and then the data is normalized. In an example, the sample data in the full-process matrix can be normalized to the (-1, 1) interval. Then, the full-process matrix after data normalization is processed by the combined feature construction method, and finally the processed full-process matrix is ​​divided into a training set and a test set according to the data set segmentation method.

[0079] In an embodiment of the present application, performing outlier elimination processing on the data of the shuffled full-process matrix to obtain data after outlier elimination processing may include:

[0080] The historical data involving shutdowns and abnormal production conditions are eliminated from the shuffled full-process matrix data.

[0081] In the embodiment of the present application, data normalization is performed on the data after outlier removal to satisfy formula (1):

[0082]

[0083] Among them, j represents the j-th dimension feature column, χ j is the sample data of the j-th dimension feature, is the sample mean of the j-th dimension feature, χ j,max is the maximum value of the j-th dimension feature, χ j,min is the minimum value of the j-th dimension feature.

[0084] In an embodiment of the present application, processing the sample data according to the combined feature construction method to obtain the processed sample data may include:

[0085] Obtain original sample features of different dimensions of sample data;

[0086] Process calculations are performed on the original sample features of different dimensions to obtain the final features;

[0087] The processed sample data is obtained according to the final features.

[0088] Specifically, after data normalization, the original sample features of different dimensions of the sample data are obtained, and then the original sample features of different dimensions are subjected to process calculation based on the knowledge of the refining process flow, thereby obtaining new process feature variables, and the new process feature variables are used as the final features to obtain the processed sample data. In one example, taking the hydrocracking process flow as an example, the original sample features of different dimensions of the sample data are subjected to process calculation based on the knowledge of the hydrocracking process, and new process feature variables are obtained as the final features; wherein the specific process calculation features may include but are not limited to the catalytic diesel blending ratio, liquid hourly space velocity, and reflux ratio.

[0089] In an embodiment of the present application, dividing the processed sample data into a training set and a test set according to a data set segmentation method may include:

[0090] Perform working condition analysis on sample data to obtain different working condition data sets;

[0091] The working condition data set is divided into a training set and a test set.

[0092] Specifically, the sample data of the processed full-process matrix is ​​subjected to working condition analysis, and the sample data under the same working condition is divided into the same data set, thereby obtaining multiple different working condition data sets, and then the working condition data sets are divided into training sets and test sets according to the set conditions. In one example, 70% of the data can be extracted from different working condition data sets as training sets, and the rest as test sets.

[0093] In an embodiment of the present application, the same layer of the neural network in the multi-scale convolutional neural network model can be dilated and convolved using convolution kernels of different sizes and dilated convolution kernels of different dilation rates.

[0094] Specifically, a multi-scale convolutional neural network is proposed based on the extended-Inception structural block. When performing multi-scale learning and extraction of feature relationships between variables in the refining process, the same layer of the full process matrix can be dilated by using convolution kernels of different sizes and dilated convolution kernels of different dilation rates. For example, 3×3 convolution, 5×5 convolution and 4×4 dilated convolution can be performed on the full process matrix in the same convolution layer. In addition, before the convolution calculation, for convolution kernels of different sizes, the input sample matrix is ​​filled with all 0s to output the same output feature map size, so that the output feature maps of the same size can be superimposed on the channel dimension as the input of the next layer. By encapsulating the multi-scale convolution kernel into a similar Inception structural block, it is beneficial to improve the efficiency of model training and reasoning.

[0095] Figure 2 Schematic diagram of the GMM-t-SNE feature evaluation framework according to an embodiment of the present application. In the embodiment of the present application, step 105, through mixed Gaussian model and t-distribution-random neighbor embedding manifold learning, determining the rationality of feature extraction of the multi-scale convolutional neural network model according to the test set and the intermediate feature map may include:

[0096] Cluster the predictor variables in the test set according to the Gaussian mixture model to obtain the corresponding Gaussian mixture model labels;

[0097] The mixed Gaussian model labels are concatenated with the test set and the intermediate feature map to obtain the data to be reduced in dimension;

[0098] According to the t-distribution-random neighbor embedding, the dimensionality reduction data is visualized to obtain two-dimensional feature clusters;

[0099] The rationality of feature extraction of multi-scale convolutional neural network models is determined based on two-dimensional feature clusters.

[0100] Specifically, the present application combines the use of mixed Gaussian models and the use of mixed Gaussian models (GMM) and t-distribution-random neighbor embedding (t-SNE) manifold learning to propose a novel feature evaluation framework GMM-t-SNE. First, based on the mixed Gaussian model, the predicted variables (labels) in the test set are clustered and labeled to obtain the corresponding GMM labels, and then the GMM labels are spliced ​​with the test set and the obtained intermediate feature map to form the data to be reduced in dimension; then the t-distribution-random neighbor embedding is used to visualize the reduced dimension data to obtain a two-dimensional feature cluster; wherein, any layer of the fully connected layer of the multi-scale convolutional neural network can be visualized in terms of reduced dimension. Finally, the rationality of the feature extraction of the multi-scale convolutional neural network model is determined based on the two-dimensional feature clusters. If the obtained two-dimensional feature clusters are consistent with the distribution of the GMM clustering conditions (i.e., GMM labels) of the corresponding predicted variables, which is manifested as only containing a certain identical GMM label in the same two-dimensional cluster, it means that MSCNN has learned reasonable discriminant features, which is helpful to predict the corresponding target variable. That is to say, the rationality of using multi-scale convolutional neural network to extract features of refining process flow is verified.

[0101] In a specific embodiment of the present application, the rationality evaluation of feature extraction of a hydrocracking process is taken as an example.

[0102] The coal diesel hydrocracking unit of a petrochemical enterprise uses straight-run diesel, catalytic diesel (FCC diesel) and reformed hydrogen (Make-up H2) as raw materials, adopts a double-agent series and one-pass hydrogenation process to produce chemical naphtha, aviation kerosene and high-quality diesel. The specific process flow is as follows: diesel is fed in proportion and mixed with hydrogen. After preheating, it first enters the hydrofining reactor (R101) to remove impurities such as sulfur, nitrogen, and metals, and then enters the cracking reactor (R102) to selectively crack diesel molecules into light hydrocarbons through hydrocracking reactions. The reactor effluent enters the fractionation process after preliminary product separation in the high and low pressure separators. The hydrogen sulfide stripping tower (C201) removes the acid gas and C4 light hydrocarbons in the product, and then obtains acid gas and liquefied gas (LPG) through the absorption and desorption tower (C205) and the naphtha stabilization tower (C206), respectively. The C201 bottom logistics after removing the light hydrocarbons enters the main distillation tower (C202) for main product separation, wherein crude naphtha is distilled from the top of the tower and passed through the naphtha distillation tower (C204) to obtain light and heavy naphtha, and the kerosene component distilled from the side line of C202 is stripped by the stripping tower (C203) to obtain aviation kerosene, and high-quality diesel products are extracted from the bottom of C202. Therefore, the hydrocracking process is highly complex, and there is process coupling within and between each local process.

[0103] Figure 3 FIG. 1 is a flow chart of a method for determining the rationality of feature extraction of a hydrocracking process according to an embodiment of the present application. Figure 3 As shown, in this specific embodiment, the method for determining the rationality of the hydrocracking process feature extraction includes the following steps.

[0104] Step 401: Process sample data of the hydrocracking process flow.

[0105] Figure 4 This is a schematic diagram of a matrix design based on hydrocracking process knowledge according to a specific embodiment of the present application. 126 dimensional process variables in the historical sample data of the hydrocracking process are selected as the initial input sample data, and the hydrorefining catalyst temperature and new hydrogen consumption are used as prediction variables. A total of 4360 sets of data were collected from the historical operation period of the device from June 2019 to September 2020. Figure 4 As shown in the figure, the 4360 sets of collected data are reshaped to obtain a 9×14 initial process matrix. Since the full process matrix fully accommodates the unit operations of each local sub-process of hydrocracking and contains the spatial relationship of all key variables in the hydrocracking process from feed to product distillation line, it is suitable for feature learning and extraction through convolution operations, thereby ultimately realizing the integration of process knowledge and deep learning.

[0106] Considering Figure 4 In the initial full-process matrix shown, non-adjacent rows (local sub-processes) have long-distance process coupling relationships. For example, the feed composition affects the product distribution, and the product distribution also reflects the temperature of the hydrocracking bed. In order to further extract the long-distance characteristic relationship between the local sub-processes, the initial full-process matrix is ​​randomly shuffled to obtain the shuffled full-process matrix.

[0107] Figure 5 FIG. 1 is a schematic diagram of data enhancement according to a specific embodiment of the present application. Figure 5 As shown, the original sample size is 4360. Figure 5 After two rounds of random row rearrangement operations were performed on the initial full-process matrix one by one, the sample capacity was expanded to 13080, achieving 3 times data enhancement. Specifically, the distribution of 126 dimensions of data in the hydrocracking process is shown in Table 1.

[0108] Table 1 Distribution of data samples in the hydrogenation process sub-process block

[0109]

[0110]

[0111] Next, the sample data in the shuffled full-process matrix are processed for outlier removal, that is, the data involving shutdowns and abnormal production conditions in the sample data of the shuffled full-process matrix are removed; then the data after outlier removal are processed for data normalization and combined feature construction, and finally, the training set and the test set are divided according to the data set segmentation method, specifically, the training set and the test set are divided according to the ratio of 80%:20%.

[0112] Step 402: Extract features of the hydrocracking process flow based on MSCNN.

[0113] First, the parameters of the MSCNN model are configured. In addition, this specific embodiment also sets up a control group based on CNN and BPNN models. The parameter configurations of the MSCNN model and the control group CNN model and BPNN model are shown in Table 2. Both MSCNN and traditional CNN have 6 layers (layer 1 to layer 6), of which the first two layers are convolutional layers (conv, the number and size of convolution kernels are indicated in Table 2). The convolution output feature map is flattened into a one-dimensional vector through the third layer and passes through the fourth and fifth fully connected layers (dense, the number of neurons is indicated in Table 2), and finally the regression prediction result is output by the sixth layer. Unlike traditional CNN, the first layer of MSCNN uses the extended-Inception structure block to realize multi-scale convolution. There are 5×5, 7×7, 4×4, and 3×34 convolution kernels of different sizes, and the fourth type of 3×3 convolution kernel uses dilated convolution (dr=2) to further enhance the multi-scale feature extraction capability. As a control, BPNN also provides dense layers with the same total number of layers (conv layers are replaced by dense layers). The network structure of the last three layers is consistent with MSCNN and CNN, and the first two layers correspond to the convolutional layers of MSCNN and CNN, each containing 128 neurons. BPNN does not contain flattening.

[0114] Table 2 MSCNN model parameter configuration

[0115]

[0116]

[0117] Next, each model was built and trained based on the tensorflow deep learning framework. The training configuration used the following parameters: the default learning rate was 0.001, the ReLU nonlinear activation function was used, the mean square error (mse) was used as the regression training metric, and the Adam optimizer with excellent performance was selected to perform gradient descent on the model. The model evaluation indicators on the test set were the root mean square error (rmse) and the determination coefficient (r 2 ), the rmse is obtained by satisfying formula (2), r 2 The solution of satisfies formula (3).

[0118]

[0119]

[0120] Among them, N is the test set capacity, y i is the true value of the predictor variable, is the model prediction value, is the mean of the true value, rmse represents the average deviation between the model prediction value and the true value, the smaller the rmse value, the more accurate the model prediction. 2 It indicates the degree to which the model can explain the overall variance of the predictor variable, r 2 Larger values ​​indicate a better model fit.

[0121] Figure 6(a) to Figure 6(d) Schematic diagram of GMM cluster analysis according to a specific embodiment of the present application. Based on GMM, the hydrofining catalyst temperature (HCT) and the fresh hydrogen consumption (H2FR) are clustered, and the BIC criterion is applied to select the GMM Gaussian component number k. The results are as follows: Figure 6(a) to Figure 6(d) As shown in the figure, the four rectangles corresponding to each GMM Gaussian component number K value in the bar graphs of Figure 6(a) and Figure 6(c) are spherical, tied, diag, and full from left to right, representing different Gaussian covariance types. The actual sample distribution affects the final clustering result of GMM. As can be seen from Figure 6(a) and Figure 6(c), when k=4 is selected, except for tied

[0122] Except for the covariance type, spherical, diag, and full all achieve the same minimum BIC score. Figure 6(b) and Figure 6(d) show the probability distribution fitting of HCT and H2FR sample data by GMM when k is 4. The left axis represents the frequency distribution of the real sample, and the right axis represents the GMM probability density function (PDF). Figure 6(b) and Figure 6(d) show that the four Gaussian components The GMM of HCT and H2FR fits the probability distribution well, which is consistent with the real frequency. represent the mean and standard deviation of each Gaussian component, respectively.

[0123] Figure 7(a) to Figure 7(h) Schematic diagram of hydrocracking process variable prediction according to a specific embodiment of the present application. HCT and H2FR were predicted on the test set based on MSCNN and compared with CNN and BPNN. The results are as follows Figure 7(a) to Figure 7(d) , where line M is the predicted value of each model, line N is the true value, and line e is the absolute error between the predicted value and the true value. Figure 7(a) to Figure 7(d) is the absolute error (ae) between the predicted value of HCT and the true value as well as the model prediction. Figure 7(a) shows the prediction of HCT based on the MSCNN-shuffled model, with rmse=0.75 and r2=0.96. Figure 7(b) shows the prediction of HCT based on the MSCNN model, with rmse=1.06 and r2=0.92. Figure 7(c) shows the prediction of HCT based on the CNN model, with rmse=1.44 and r2=0.86. Figure 7(d) shows the prediction of HCT based on the BPNN model, with rmse=1.65 and r2=0.81. Figure 7(e) to Figure 7(h) The corresponding H2FR prediction results are shown in Figure 7(e) for the prediction of H2FR based on the MSCNN-shuffled model, rmse = 1053.0, r2 = 0.99; Figure 7(f) for the prediction of H2FR based on the MSCNN model, rmse = 1476.0, r2 = 0.98; Figure 7(g) for the prediction of H2FR based on the CNN model, rmse = 1857.0, r2 = 0.97; Figure 7(h) for the prediction of H2FR based on the BPNN model, rmse = 2588.0, r2 = 0.94. Shuffled means that the training set is randomly rearranged. From Figure 7(a) and Figure 7(e), it can be seen that the rmse of MSCNN-shuffled for HCT prediction reaches a minimum of 0.75, and the r2 reaches a maximum of 0.96, and the rmse of MSCNN-shuffled for H2FR prediction reaches a minimum of 1053, and the r2 reaches a maximum of 0.99, and the prediction results are all optimal. In Figure 7(b) and Figure 7(f), the rmse of MSCNN prediction for HCT and H2FR are 1.06 and 1476, and the r2 is 0.92 and 0.98, respectively, both of which are suboptimal. Figure 7(a) to Figure 7(h) The prediction results of CNN and BPNN are relatively poor, both inferior to MSCNN-shuffled and MSCNN, indicating that the MSCNN algorithm based on knowledge fusion fully extracts the characteristic relationships of different scales among the local process flows of hydrocracking and significantly improves the prediction performance of the model.

[0124] Step 403: Perform feature rationality evaluation based on the GMM-t-SNE feature evaluation framework.

[0125] Figure 8(a) to Figure 8(j) Schematic diagram of t-SNE visualization of IFM according to a specific embodiment of the present application. Each model outputs IFM for HCT and H2FR predictions at the same time, and the IFM is concatenated with the four different GMM labels of the predicted target HCT and H2FR, and the GMM labels are circled with different shapes. Figure 2 The process uses t-SNE to reduce the dimension of IFM to two dimensions (dimension 1, dimension 2). In order to obtain a good visualization effect, the perplexity parameter of the t-SNE algorithm is set to 100. The final visualization result is as follows: Figure 8(a) to Figure 8(j) shown.

[0126] The feature visualization results of different models for HCT and H2FR prediction are as follows: Figure 8(a) to Figure 8(e) and Figure 8(f) to Figure 8(j)In order to compare the rationality of the features extracted by different model algorithms, Figure 8(a) and Figure 8(f) respectively show the dimensionality reduction results of the original sample matrices of HCT and H2FR as blank controls, Figure 8(b) is the prediction of HCT based on the BPNN model, Figure 8(c) is the prediction of HCT based on the CNN model, Figure 8(d) is the prediction of HCT based on the MSCNN model, Figure 8(e) is the prediction of HCT based on the MSCNN-shuffled model, Figure 8(g) is the prediction of H2FR based on the BPNN model, Figure 8(h) is the prediction of H2FR based on the CNN model, Figure 8(i) is the prediction of H2FR based on the MSCNN model, and Figure 8(j) is the prediction of H2FR based on the MSCNN-shuffled model. As shown in the blank control Figure 8(a), the two-dimensional cluster distribution obtained by t-SNE dimensionality reduction of the original sample cannot correspond to the four GMM labels of HCT one by one. Instead, the GMM label of the same HCT appears in multiple clusters, and the same cluster also contains GMM labels of different HCTs, which indicates that the data form of the original sample is not conducive to the prediction of HCT. The blank control of H2FR in Figure 8(f) also shows the same result. The MSCNN algorithm based on knowledge fusion fully extracts the feature relationship of the local-global spatial domain of the hydrocracking process through multi-scale convolution. By observing the t-SNE two-dimensional clusters of these features in Figure 8(d), it can be seen that the features under the same GMM label of HCT form independent clusters, and the features under different GMM labels belong to different clusters. This shows that the features extracted by MSCNN are closely related to the predicted target HCT, and the distribution of the features is consistent with the distribution of GMM labels of HCT, which can significantly improve the prediction accuracy of HCT. Figure 8(e) The MSCNN-shuffled training set has been enhanced with data, and more abundant feature relationships have been extracted. Therefore, the feature distribution is optimally consistent with the GMM labels distribution of HCT, indicating that the features extracted by MSCNN-shuffled are the most reasonable. The CNN shown in Figure 8(c) does not use multi-scale convolution, and its feature extraction capability is insufficient, resulting in poor distribution consistency. The BPNN in Figure 8(b) is a fully connected network, and the extracted features only come from the global nonlinear transformation of the original samples. It can be seen from the feature distribution that its rationality is weaker than that of MSCNN and CNN, and is only slightly better than that of the original samples. Figure 8(g) to Figure 8(j) The characteristic evaluation results of H2FR shown are the same as those of HCT.

[0127] Fig. 9 FIG. 1 is a structural block diagram of a device for determining the rationality of refining process feature extraction according to an embodiment of the present application. Fig. 9 As shown, the embodiment of the present application provides a device for determining the rationality of refining process feature extraction, which may include:

[0128] Memory 910, configured to store instructions; and

[0129] The processor 920 is configured to call instructions from the memory and implement any of the above methods for determining the rationality of refining process flow feature extraction when executing the instructions.

[0130] Specifically, in the embodiment of the present application, the processor 920 may be configured to:

[0131] Obtain industrial historical sample data;

[0132] Construct a full-process matrix based on industrial historical sample data to obtain training sets and test sets;

[0133] Training a multi-scale convolutional neural network model according to the training set to obtain a trained multi-scale convolutional neural network model;

[0134] Input the test set into the trained multi-scale convolutional neural network model to obtain the intermediate feature map;

[0135] Through mixed Gaussian model and t-distribution-random neighbor embedding manifold learning, the rationality of feature extraction of multi-scale convolutional neural network model is determined based on test sets and intermediate feature maps.

[0136] Furthermore, the processor 920 may be further configured to:

[0137] The full process matrix is ​​constructed based on the industrial historical sample data to obtain the training set and test set including:

[0138] The initial full-process matrix is ​​obtained based on industrial historical sample data;

[0139] Randomly rearrange and shuffle the initial full-process matrix to obtain a shuffled full-process matrix;

[0140] The shuffled full-process matrix is ​​processed to obtain the training set and the test set.

[0141] Furthermore, the processor 920 may be further configured to:

[0142] The shuffled full-process matrix is ​​processed to obtain the training set and test set including:

[0143] Performing outlier elimination processing on the data of the shuffled full-process matrix to obtain outlier-eliminated data;

[0144] Normalize the data after outlier removal to obtain sample data;

[0145] Processing the sample data according to the combined feature construction method to obtain processed sample data;

[0146] The processed sample data is divided into training set and test set according to the data set segmentation method.

[0147] Furthermore, the processor 920 may be further configured to:

[0148] The outlier removal process is performed on the data of the shuffled full-process matrix to obtain the data after outlier removal process including:

[0149] The historical data involving shutdowns and abnormal production conditions are eliminated from the shuffled full-process matrix data.

[0150] Furthermore, the processor 920 may be further configured to:

[0151] The data after outlier removal is normalized to meet formula (1):

[0152]

[0153] Among them, j represents the j-th dimension feature column, χ j is the sample data of the j-th dimension feature, is the sample mean of the j-th dimension feature, χ j,max is the maximum value of the j-th dimension feature, χ j,min is the minimum value of the j-th dimension feature.

[0154] Furthermore, the processor 920 may be further configured to:

[0155] The sample data is processed according to the combined feature construction method to obtain the processed sample data including:

[0156] Obtain original sample features of different dimensions of sample data;

[0157] Process calculations are performed on the original sample features of different dimensions to obtain the final features;

[0158] The processed sample data is obtained according to the final features.

[0159] Furthermore, the processor 920 may be further configured to:

[0160] According to the data set segmentation method, the processed sample data is divided into a training set and a test set, including:

[0161] Perform working condition analysis on sample data to obtain different working condition data sets;

[0162] The working condition data set is divided into a training set and a test set.

[0163] In an embodiment of the present application, the same layer of the neural network in the multi-scale convolutional neural network model performs dilated convolution through convolution kernels of different sizes and dilated convolution kernels of different dilation rates.

[0164] Furthermore, the processor 920 may be further configured to:

[0165] Through mixed Gaussian model and t-distribution-random neighbor embedding manifold learning, the rationality of feature extraction of multi-scale convolutional neural network model is determined based on test set and intermediate feature map, including:

[0166] Cluster the predictor variables in the test set according to the Gaussian mixture model to obtain the corresponding Gaussian mixture model labels;

[0167] The mixed Gaussian model labels are concatenated with the test set and the intermediate feature map to obtain the data to be reduced in dimension;

[0168] According to the t-distribution-random neighbor embedding, the dimensionality reduction data is visualized to obtain two-dimensional feature clusters;

[0169] The rationality of feature extraction of multi-scale convolutional neural network models is determined based on two-dimensional feature clusters.

[0170] Through the above technical scheme, first obtain industrial historical sample data, then build a full process matrix based on the industrial historical sample data to obtain a training set and a test set; train a multi-scale convolutional neural network model based on the training set to obtain a trained multi-scale convolutional neural network model; further input the test set into the trained multi-scale convolutional neural network model to obtain an intermediate feature map; finally, through a mixed Gaussian model and t-distribution-random neighbor embedding manifold learning, according to the test set and the intermediate feature map, determine the rationality of the feature extraction of the multi-scale convolutional neural network model. This application extracts features from the refining process flow through a multi-scale convolutional neural network, and evaluates the rationality of the features extracted by the multi-scale convolutional neural network based on a mixed Gaussian model and t-distribution-random neighbor embedding popular learning, thereby verifying the effectiveness of the multi-scale convolutional neural network in extracting features from the refining process flow, which is beneficial to solving the problem in the prior art that the feature extraction of the process flow cannot be achieved through the fusion of multi-scale convolutional networks and process flow knowledge.

[0171] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0172] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0173] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0174] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0175] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0176] The memory may include non-permanent memory in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.

[0177] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0178] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0179] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included within the scope of the claims of the present application.

Claims

1. A method for determining the rationality of feature extraction of a refining process flow, It is characterized in that include: Obtaining sample data of process variables in historical sample data of the process flow as initial input; Constructing a full-process matrix according to the sample data of the initial input to obtain a training set and a test set; Training a multi-scale convolutional neural network model according to the training set to obtain a trained multi-scale convolutional neural network model; Inputting the test set into the trained multi-scale convolutional neural network model to obtain an intermediate feature map; Determining the rationality of feature extraction of the multi-scale convolutional neural network model according to the test set and the intermediate feature map through mixed Gaussian model and t-distribution-random neighbor embedding manifold learning; Wherein, the method of learning the mixed Gaussian model and the t-distribution-random neighbor embedding manifold and determining the rationality of feature extraction of the multi-scale convolutional neural network model according to the test set and the intermediate feature map includes: Clustering the predicted variables in the test set according to the Gaussian mixture model to obtain the corresponding Gaussian mixture model label GMM label; The mixed Gaussian model label is concatenated with the test set and the intermediate feature map to obtain the data to be reduced in dimension; Performing dimension reduction visualization on the data to be dimensionally reduced according to t-distribution-random neighbor embedding to obtain two-dimensional feature clusters; Determining the rationality of feature extraction of the multi-scale convolutional neural network model according to the two-dimensional feature clusters; Among them, if the obtained two-dimensional feature cluster is consistent with the GMM label distribution of the corresponding prediction variable, which is manifested as the same two-dimensional cluster containing only a certain same GMM label, it indicates that the multi-scale convolutional neural network model has learned reasonable discriminant features.

2. The method according to claim 1, It is characterized in that The step of constructing a full-process matrix according to the sample data initially input to obtain a training set and a test set includes: Obtaining an initial full-process matrix according to the sample data of the initial input; Randomly rearrange and shuffle the initial full-process matrix to obtain a shuffled full-process matrix; The shuffled full-process matrix is ​​processed to obtain a training set and a test set.

3. The method according to claim 2, It is characterized in that The processing of the shuffled full-process matrix to obtain a training set and a test set includes: Performing outlier elimination processing on the data of the shuffled full-process matrix to obtain outlier-eliminated data; Normalizing the data after the outlier removal process to obtain sample data; Processing the sample data according to the combined feature construction method to obtain processed sample data; The processed sample data is divided into a training set and a test set according to a data set segmentation method.

4. The method according to claim 3, It is characterized in that The step of performing outlier elimination processing on the shuffled full-process matrix data to obtain outlier-eliminated data includes: The historical data involving shutdown and abnormal production conditions in the shuffled full-process matrix data are eliminated.

5. The method according to claim 3, It is characterized in that The data normalization after the outlier removal process satisfies the following formula: Among them, j represents the j-th dimension feature column, x j is the sample data of the j-th dimension feature, is the sample mean of the j-th dimension feature, x j,max is the maximum value of the j-th dimension feature, x j,min is the minimum value of the j-th dimension feature.

6. The method according to claim 3, It is characterized in that The processing of the sample data according to the combined feature construction method to obtain processed sample data includes: Obtaining original sample features of different dimensions of the sample data; Performing process calculation on the original sample features of different dimensions to obtain final features; The processed sample data is obtained according to the final features.

7. The method according to claim 3, It is characterized in that The step of dividing the processed sample data into a training set and a test set according to the data set segmentation method comprises: Performing working condition analysis on the sample data to obtain different working condition data sets; The operating condition data set is divided into a training set and a test set.

8. The method according to claim 1, It is characterized in that The same layer of the neural network in the multi-scale convolutional neural network model performs dilated convolution through convolution kernels of different sizes and dilated convolution kernels of different dilation rates.

9. A device for determining the rationality of feature extraction of a refining process flow, It is characterized in that include: a memory configured to store instructions; as well as A processor is configured to call the instructions from the memory and implement the method for determining the rationality of refining process feature extraction according to any one of claims 1 to 8 when executing the instructions.