Root cause analysis of chemical process and yield prediction modeling method using same

Abstract

According to one embodiment of the invention, a root cause analysis and yield prediction method is disclosed. The method for analyzing the root cause of the yield change of the chemical process and predicting the yield comprises the following steps of: preprocessing data for predicting the yield; and a step of analyzing a root cause by using the preprocessed yield prediction data, wherein the root cause analysis step comprises a step of extracting a process key factor and a step of analyzing the process influence degree of the extracted key factor. Also, one embodiment of the present invention discloses a computer-readable recording medium on which a computer program is recorded for implementing the root cause analysis and yield prediction method.

Description

Technical Field

[0001] The present invention relates to a root cause analysis of a chemical process, and more specifically, to a yield prediction system and method for screening out key factors that promote yield changes and applying them to a yield prediction model. Background Art

[0002] When implementing a series of processes consisting of multiple steps, it is very important to ensure the integrity and reliability of the organic connection of each process. In order to achieve this integrity, it is necessary to develop an effective process management system so that the input values ​​of the main equipment in the process can be used to determine whether each process has abnormalities and diagnose the causes.

[0003] Most of the historical data generated by industrial processes usually have a small number of variables and a linear data structure, so existing algorithms can also obtain sufficient prediction / classification results. However, due to the development of ICT and sensor technology, the field of manufacturing processes has begun to generate data with hundreds or thousands of variables. In particular, modern industrial processes, such as chemical and manufacturing processes, and power plants, have made various efforts to reduce costs and pursue maximum profits while meeting safety, health and environmental regulations, which has led to an increasing size and complexity of data.

[0004] Therefore, it is extremely important to filter out and manage the data that has a greater impact on profit creation from the above complex and large-scale data. For example, in commercial chemical processes, process operating conditions have an impact on catalyst activity in the short / long term, and catalyst activity is directly related to product output. Therefore, it is extremely important to predict catalyst activity from a short / long-term perspective.

[0005] From a short-term perspective, it is necessary to understand the changes in catalyst activity according to process operating conditions, and to increase the output of products by optimizing the operating conditions to improve catalyst activity. From a long-term perspective, in commercial chemical processes using catalysts, the longer the process runs, the less active the catalyst becomes, resulting in reduced activity, so it is necessary to replace it after a period of time. In this way, replacing the catalyst is time-consuming and expensive, so it is important to predict the future catalyst activity and determine the life / replacement timing of the catalyst.

[0006] Therefore, it is necessary to develop new technologies in order to understand the process operating conditions (key factors) that have a great impact on catalyst activity and reflect them, thereby improving the prediction accuracy of catalyst reaction activity.

[0007]

Prior technical literature

[0008] [Patent Literature]

[0009] (Patent Document 1) Korean Patent Publication No. 2018-0131246 (20181210)

[0010] (Patent Document 2) Korean Patent Publication No. 2015-0018681 (20150224)

[0011] (Patent Document 3) Japanese Patent Publication No. 2020-166749 (20201008)

[0012] (Patent Document 4) Korean Patent Publication No. 10-2019-0060547 (20190603)

[0013] (Patent Document 5) Japanese Patent Publication No. 2022-520643 (20220331)

[0014] (Patent Document 6) Korean Registered Patent Gazette No. 10-2218287 (20210222) Summary of the invention

[0015] Issues to be solved

[0016] The purpose of the present invention is to provide a catalyst activity prediction system and method that reflects key factors in commercial chemical processes, so as to solve the problems existing in the process change cause analysis technology in traditional technologies, screen out the key factors that promote yield changes, and further apply them to the prediction model, thereby improving the prediction accuracy of catalyst activity changes.

[0017] The purpose of the present invention is to provide a catalyst activity prediction system and method that reflects key factors in commercial chemical processes, so as to determine product sales plans and catalyst replacement timing through accurate catalyst activity prediction, and effectively manage the time and cost required for product production and catalyst replacement.

[0018] Another object of the present invention is not limited to the above-mentioned object, and a person skilled in the art can clearly understand another object not mentioned from the following content.

[0019] Problem Solution

[0020] One embodiment of the present invention discloses a root cause analysis and yield prediction method, which is a root cause analysis and yield prediction method for yield variation in a chemical process.

[0021] include:

[0022] The step of preprocessing the data for yield prediction; and

[0023] The step of analyzing the root cause using the yield prediction data of the pre-treatment,

[0024] The root cause analysis step includes the steps of extracting process key factors and analyzing the process impact of the extracted key factors.

[0025] One embodiment of the present invention discloses a computer-readable recording medium having a computer program recorded thereon, for implementing the root cause analysis and yield prediction method.

[0026] Effects of the Invention

[0027] The present invention can use XAI to extract key factors, display the contribution of each key factor, and further analyze the influence of each key factor in detail. That is, the user can confirm the contribution of different labels of key factors on the user interface (UI) screen, confirm the labels that have an impact on the labels on a specific date according to the contribution order, and output the actual observation data of the date, and verify the key factor contribution results according to the XAI analysis, thereby improving the reliability of the XAI analysis.

[0028] Furthermore, the present invention can screen out key factors according to process operating conditions and apply them to prediction models, thereby improving the accuracy of predictions of catalyst activity changes. Through accurate catalyst activity predictions, the sales plan of the product and the timing of catalyst replacement can be determined, and the time and cost required for product production and catalyst replacement can be effectively managed. Changes in catalyst activity can be grasped according to process operating conditions, and catalyst activity can be improved and product output can be increased by optimizing operating conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 is a block diagram of a root cause analysis and yield prediction system according to an embodiment of the present invention.

[0030] Figure 2 It is a flow chart of the root cause analysis and yield prediction method described in one embodiment.

[0031] Figure 3 It is a schematic diagram of a data preprocessing method according to an embodiment.

[0032] Figure 4 The flowchart of a method for visualizing analysis results according to an embodiment of the present invention.

[0033] Figures 5 to 9 It is a user interface (UI) view of an embodiment outputting analysis results.

[0034] Fig.10 It is a schematic flow chart of a segment analysis method described in an embodiment.

[0035] Figures 11 to 16 Schematic diagram of a segment analysis process according to an embodiment.

[0036] Fig.17 and Fig.18 It is a schematic diagram of a data realization method according to an embodiment.

[0037] Figures 19 to 21 FIG. 1 is a schematic diagram of a method for reflecting catalyst aging factors according to an embodiment of the present invention.

[0038] Fig. 22 It is a view of the yield prediction result described in an embodiment.

[0039] * Reference Numbers *

[0040] 100: Root cause analysis and yield prediction system 110: Data preprocessing unit

[0041] 120: Root Cause Analysis Department 130: Yield Prediction Department

[0042] 140: Segment Analysis Unit 150: Data Realization Processing Unit

[0043] 160: Aging factor analysis unit 200: Data storage unit DETAILED DESCRIPTION

[0044] [Preferred form of the embodiment of the invention]

[0045] One embodiment of the present invention discloses a root cause analysis and yield prediction method, which is a root cause analysis and yield prediction method for yield variation in a chemical process.

[0046] include:

[0047] The step of preprocessing the data for yield prediction; and

[0048] The step of analyzing the root cause using the yield prediction data of the pre-treatment,

[0049] The root cause analysis step includes the steps of extracting process key factors and analyzing the process impact of the extracted key factors. [Specific implementation method]

[0051] The above-mentioned several objects, other several objects, several features and several advantages of the present invention can be easily understood through the following preferred embodiments related to the accompanying drawings. However, the present invention is not limited to the embodiments described here, but is specifically implemented through other forms. On the contrary, the embodiments introduced here are provided to make the disclosed content comprehensive and complete, and to fully convey the ideas of the present invention to those skilled in the art.

[0052] In this specification, when the terms first, second, etc. are used to describe components, the components should not be limited to the terms. The terms are only used to distinguish a component from other components. The embodiments described and illustrated herein include their complementary embodiments.

[0053] Unless otherwise specifically mentioned in the context of this specification, the singular form also includes the plural form. In the expressions used in the specification, "comprising", "consisting of" and "presenting" do not exclude the existence or addition of one or more other components other than the mentioned components.

[0054] In this specification, the term "software" means technology for running hardware in a computer, the term "hardware" means a device or machine that constitutes a computer (CPU, memory, input device, output device, peripheral device, etc.), the term "step" means a series of processing or operations connected in time series in order to achieve a given purpose, the term "computer program", "program" or "algorithm" means an instruction set suitable for computer processing, and the term "program recording medium" means a computer-readable recording medium that sets a program and executes or records a stream of general programs.

[0055] In this specification, terms such as "part", "module", "unit", "block", "board", etc. used to indicate components of the invention may mean a physical, functional or logical unit for processing at least one function or action, which may be presented as more than one hardware or software or firmware, or may be presented as a combination of more than one hardware, software and / or firmware.

[0056] In this specification, "processing device", "computer", "computing device", "server device", "server" can be presented as a system with an operating system such as Windows, Mac, Linux, a computer program, a memory, a number of application programs and a memory device (for example, HDD, SSD). For example, the computer can be a desktop computer or a laptop computer, a mobile terminal and other devices, but it is illustrative and not limited to this. The mobile terminal can be one of the mobile wireless communication devices, such as a smart phone, a tablet computer or a PDA.

[0057] Hereinafter, the present invention will be described in detail with reference to the accompanying drawings. Various specific contents written in the process of describing the following specific embodiments are used to describe in more detail and help understand the invention. However, it can be understood that those skilled in the art of the present invention should be able to know that it can be adopted even without these multiple specific contents. And it is forewarned that in order to prevent confusion when explaining the invention, parts that belong to the well-known or well-known conventional technology and have little to do with the invention will not be described.

[0058] Figure 11 is a schematic block diagram of a root cause analysis and yield prediction system according to an embodiment of the present invention. In the following description of this specification, it is assumed that the root cause analysis and yield prediction system according to the present invention (hereinafter referred to as "yield prediction system") is applied to an olefin production process for explanation. For example, the yield prediction system according to the present invention can be applied to a PDH (Propane DeHydration) process for preparing propylene using propane as a raw material, through which hydrogen can be extracted from propane to generate one of the olefins, namely, propylene.

[0059] In one embodiment of the present invention, the yield prediction system can predict the yield during the second period based on the process operation data collected during the first period. In this case, the first and second periods can be periods of the same length or periods of different lengths, for example, they can be set to 4 years. Preferably, a period can be related to the life of the catalyst used in the chemical process, for example, when the life of the catalyst is 4 years, a period can be set to 4 years.

[0060] like Figure 1 As shown, the yield prediction system 100 of one embodiment may include a data preprocessing unit 110, a root cause analysis unit 120 and a yield prediction unit 130. Each of these components (110 to 140) may be implemented as programmable software that can be run in a computer device, and may be implemented in combination with firmware and hardware implementation parts as needed. In an alternative embodiment, the yield prediction system 100 of the present invention further includes a section analysis unit 140, a data realization processing unit 150 and an aging factor analysis unit 160.

[0061] The data preprocessing unit 110 is a functional unit that collects and extracts data from the data storage unit 200 and performs preprocessing. Figure 3 The operation of the data preprocessing unit 110 will be described.

[0062] The root cause analysis unit 120 uses the data preprocessed by the data preprocessing unit 110 to extract the key factors of the process, and selects the influence of the selected key factors on a daily basis according to the needs of the user, analyzes the root causes, and further analyzes the impact on the label (yield, selectivity, conversion rate, etc.). The root cause analysis unit 120 may include a machine learning algorithm to extract the key factors using the preprocessed data. Figures 4 to 9 The exemplary operation of the root cause analysis unit 120 will be described.

[0063] The yield prediction unit 130 is a functional unit that uses the preprocessed data and the extracted key factors to predict the process yield. In one embodiment, the yield prediction unit 130 includes a learning algorithm based on machine learning, which can use the preprocessed data and key factors to learn more than one yield prediction model, and use the learned yield prediction model to output a yield prediction result.

[0064] The segment analysis unit 140 can divide a period into multiple segments based on the preprocessed data. For example, for a period (e.g., 4 years), based on the change in given factors such as process temperature and yield caused by the life (aging) of the catalyst used in the process, the period is divided into multiple periods. Figures 10 to 16 The schematic operation of the section analysis unit 140 is described below.

[0065] The data realization processing unit 150 is a functional unit that realizes the data used in the yield prediction and generates it into a form suitable for input into the yield prediction model. The yield prediction simulation system described in the present invention uses the data of the previous cycle (the first cycle) to predict the next cycle (the second cycle). At this time, the data of the second cycle can be generated based on the data of the first cycle and input into the yield prediction model. Fig.17 and Fig.18 The schematic operation of the data actualization processing unit 130 is described.

[0066] The aging factor analysis unit 160 is a functional unit for reflecting the aging of the catalyst used in the process in order to more accurately predict the process yield. Generally, catalysts have different lifespans depending on their type, and their aging trends vary during their lifespan. In particular, as the catalyst ages faster in the latter half of its life, it may be difficult to accurately reflect it in the yield prediction model. Therefore, one embodiment of the present invention will further consider the aging factor. For example, it can be reflected in the process data as a weighted value. The process data is the aging factor of the catalyst calculated based on time, and its value is input into the yield prediction model. Through this process, the yield prediction performance can be improved. Below, refer to Figures 19 to 21 The schematic operation of the aging factor analysis unit 160 is described below.

[0067] Figure 2 It is a schematic flow chart of using the yield prediction system 100 to analyze root causes and predict yield.

[0068] like Figure 2As shown, the root cause analysis and yield prediction method described in one embodiment may include: a step of preprocessing the yield prediction data (S10), a step of analyzing the root cause using the preprocessed yield prediction data (S20), and a step of learning the yield prediction model based on the preprocessed yield prediction data and the extracted key factors (S30). At this time, the step of analyzing the root cause S20 may include: a step of extracting process key factors S21 and a step of analyzing the process influence of the extracted key factors S22. At this time, the process influence analysis step S22 visualizes and outputs the influence of the selected given label through the user interface (UI) according to the user input.

[0069] Furthermore, in an alternative embodiment, the yield prediction system 100 may further include: a segment analysis step (S40) of splitting a cycle into a plurality of segments based on preprocessed data, a data realization step (S50) of generating and realizing data for the second cycle based on data of the first cycle and the first half of data of the second cycle, and an aging factor analysis step (S60) of calculating an aging factor of a catalyst used in the process.

[0070] When further describing each step in detail, first, step (S10) pre-processes the data for yield prediction. Figure 3 FIG. 2 shows a data preprocessing method according to an embodiment. Figure 3 As shown, the data preprocessing step (S10) may include: a step of preprocessing the data for yield prediction in units of minutes (S110), a step of selecting an analysis object label (S120), a step of extracting time and daily unit data for the data of the selected label in the data preprocessed in units of minutes (S130), and a step of performing outlier processing and missing value interpolation on the daily unit data (S140).

[0071] In order to pre-process the data in step (S110), the data storage unit 200 collects and extracts the required data. At this time, for example, the data storage unit 200 can be implemented as a database, but the data form is not particularly limited. In one embodiment, the yield prediction data extracted by the data storage unit 200 may include: (i) process operation data of olefin producers, (ii) laboratory data including LIMS data, (iii) plant event data including data related to the time when the plant is not operating normally; and (iv) historical yield, conversion rate and selectivity data related to olefin production.

[0072] The process operation data of the olefin producer (PDH plant) may be sensor data collected from sensors installed in various equipment (e.g., reactors, flow paths, etc.) of the plant. For example, each sensor may be a variable detection sensor that can observe process operation conditions such as temperature, pressure, flow, and components, and data may be collected from each sensor in units of minutes.

[0073] The process operation data can be stored in the data storage unit 200 according to different sections, units and tags. At this time, the unit is a mid-size collection of tags gathered in the factory, and the section is a large-size collection of units gathered. Multiple sections are gathered together to form the entire PDH factory. In addition, the tag can serve as an identifier to identify the various sensors installed in the factory. That is, each sensor is assigned an inherent tag. In one embodiment, there can be more than 9,000 tags in the PDH factory. In the following description, unless there is a particular concern of confusion, the data output by the sensor corresponding to each tag is referred to as a "tag" or "tag data."

[0074] Laboratory data may include laboratory information management system (LIMS) data. In one embodiment, for accurate yield prediction, not only actual observation data (label data) but also laboratory data may be applied. Moreover, when label data has abnormal values ​​or missing values, laboratory data may also be used for processing, i.e., interpolation. As an alternative embodiment, laboratory data may also be omitted.

[0075] For example, plant event data may include data related to the time of non-normal operation (shut-down history), major overhaul / minor overhaul and other data, which can be used to analyze and process abnormal values ​​or missing values ​​of tag data. Historical yield values ​​include historical yield data related to olefin production. And at this time, in addition to historical yields, conversion and selectivity values ​​may be further included. In the following description of this specification, unless there is a concern of confusion, yield, conversion rate and selectivity are collectively referred to as "yield", and, from the perspective of machine learning, yield (i.e., yield, conversion rate, selectivity) is sometimes referred to as "target".

[0076] As described above, the data stored in the data storage unit 200 is continuously accumulated in units of a given set period, and can be data from the current time point to the past 5 years. At this time, the set period unit can be a second unit or a minute unit. For example, data can be collected in units of 30 seconds, and then converted into minute units and / or time units for analysis and stored. However, this set period unit is indicative and is not limited to a specific period. In addition, a period can be set in relation to the life of the catalyst, and the data accumulation period is not 5 years, but can be set to other periods.

[0077] The yield prediction data extracted from the data storage unit 200 is preprocessed into minute-based data in step S110. For example, when receiving second-based data from the data storage unit 200, it is converted into minute-based data, and when an abnormal value or missing value occurs, abnormal value processing and missing value interpolation are performed.

[0078] Next, in step (S120), the target label for analysis is selected. For example, all label data received from all sensors set up in the factory can be used to implement the following key factor analysis and yield prediction, etc., but, preferably, part of the label data in the entire label data can be screened out, and the screened label data can be used to implement subsequent (for example, after step S130) time / day unit data extraction, key factor analysis and yield prediction. At this time, the target label for analysis is selected in step (S120). For example, label selection can be based on previous research and the knowledge and experience of field technicians to select labels that are considered to be helpful for analysis.

[0079] In one embodiment, the step of selecting the analysis target label (S120) may be performed before the step of preprocessing the minute unit data (S110). In this case, the minute unit data preprocessing (S110) may be performed only on the label selected as the analysis target.

[0080] When the analysis target label is selected in step (S120), in the subsequent step (S130), the time unit data is extracted and processed again to extract the day unit data. At this time, for data integration, the factory data (label data) and LIMS data can also be used for data integration. And, as an alternative embodiment, the analysis target label selection step (S120) can also be implemented after the time unit data is extracted. At this time, after the time unit data is extracted for all process data, only the analysis target label can be extracted into day unit data.

[0081] After step (S130) extracts the daily unit data, step (S140) implements data preprocessing. For example, data preprocessing includes outlier processing and missing value interpolation. Outlier processing is to screen outliers and remove or correct them, and then only apply the sorted values ​​as valid input values. In addition, for the sections that are screened as outliers and removed or the sections that do not have process data due to factory interruption, missing value interpolation is implemented. For example, missing value interpolation is to generate new data through linear regression and distribution-based random number generation to perform interpolation.

[0082] The yield prediction data preprocessed through the above steps is sorted and converted into a learning model data format for machine learning, and then stored in the data storage unit 200 or any other storage unit.

[0083] Reference again Figure 2 It can be seen that step (S20) uses the data preprocessed as above to perform root cause analysis. In one embodiment, the root cause analysis includes: a step of extracting process key factors (S21) and a step of analyzing the process impact of the extracted key factors (S22). At this time, in the step of extracting key factors (S21), the preprocessed yield prediction data can be applied to the machine learning algorithm to extract the key factors.

[0084] Preferably, the step (S21) is applied to a describable artificial intelligence (XAI) algorithm to extract process key factors. In one embodiment, the key factor extraction step (S21) applies the SHAP algorithm to extract label values, i.e., key factors of yield, conversion rate and selectivity. For the key factors thus extracted, step (S22) analyzes the quantitative influence (contribution) of the key factors and outputs them in a visualized manner.

[0085] The Shapley Additive exPlanations (SHAP) algorithm uses Shapley Value, and the basic idea of ​​Shapley Value is as follows: in order to grasp the contribution of a specific variable to the predictive power, this method calculates the contribution of the variable while comparing the result values ​​obtained when all variable combinations related to the specific variable are input. In one embodiment of the present invention, the preprocessed data can be applied to the SHAP algorithm to extract the Shapley value of each process label, and based on this, the key factors are extracted, and the contribution is analyzed according to different key factors.

[0086] The step (S22) of analyzing the process influence degree is to analyze the influence degree (contribution degree) of a given key factor selected by the user input among the key factors extracted in the step (S21), and to visualize and output it through a user interface (UI) (e.g., a computer screen). For this, reference is made to Figures 4 to 9 Specific description.

[0087] Figure 4 The flowchart of a method for visualizing analysis results according to an embodiment of the present invention. Figures 5 to 9 shows the output Figure 4 The schematic structure of the user interface (UI) for analyzing data is shown in the flowchart.

[0088] like Figure 4 As shown in FIG. 2 , a user can set a specific period and request the XAI analysis results for the period (S210). The yield prediction system of the present invention receives the request, processes the XAI analysis results for the period, and outputs the analysis results through the user interface (UI) on the computer screen (S220). For example, when a user sets a specific period and inputs it into the system, the following is obtained: Figure 5 The resulting screen is shown. Figure 5 It is an output screen output to the same display screen as the computer screen, for example, it may include a total contribution display area 310, a tag contribution display area 320 and a daily contribution display area 330.

[0089] The total contribution display area 310 is an area that displays the contribution of the extracted key factors (labels) in the order of the largest contribution as a bar chart. The individual label contribution display area 320 is an area that displays the influence of the selected label along with the time when the user selects a specific label. The daily contribution display area 330 is an area that displays the contribution of each key factor of the selected date through a bar chart and a cumulative chart when the user selects a certain date in the individual label contribution display area 320.

[0090] In one embodiment, an analysis result output (S220) only outputs the entire contribution display area 310 to the screen, or only outputs the entire contribution display area 310 and each tag contribution display area 320, so that the user can easily and quickly confirm and analyze the main key factors. Figure 6 It is an enlarged view of the entire contribution display area 310, and the key factors are arranged in order of their greatest impact on the label (yield, conversion rate and / or selectivity) within the set period, so that the impact (contribution) of each key factor is clear at a glance.

[0091] Figure 6In the chart, the vertical axis represents the main key factors (labels), and the horizontal axis represents the contribution of each key factor with numerical values. For example, the label (T) on the far right of the vertical axis is a flow-related label, which means that the flow data detected on the flow path set at a specific location in the factory has the greatest contribution to the label value during the set period.

[0092] Users can choose to output to Figure 6 When a specific tag is selected from the tags in the screen (for example, by placing the mouse cursor on the icon of the tag and clicking it), the contribution of the selected tag is output to the tag contribution display area 320. For example, when the user selects Figure 6 When the rightmost label (T) is Figure 7 As shown, the contribution corresponding to the label (T) can be output.

[0093] like Figure 7 As shown, each tag contribution display area 320 can output individual factor analysis charts and integrated factor analysis charts. The individual factor analysis charts are displayed according to the time on the vertical axis. Figure 6 As shown in the figure, the contribution of the tag (T) selected by the user to the target is integrated into the factor analysis chart, and several other key factors are also accumulated, showing the overall contribution of each of these key factors accumulated over time. In the chart, red indicates a positive impact on the increase of the target value, and blue indicates a negative impact on the decrease of the target value.

[0094] In one embodiment, the yield prediction system of the present invention can provide Figure 7 As shown in the figure, the user can select a function for a specific date in the UI screen. For example, when the user selects Figure 7 When the date indicated by the middle arrow 30 is reached (for example, by placing the mouse cursor on the chart or the vertical axis and clicking on it, etc.) (S230 as shown in FIG4 ), the daily contribution display area 330 will be generated and output. At this time, the daily contribution display area 330 can be generated and output in the area below the tag contribution display area 320 in the existing screen, and the user can scroll on the screen to view the entire area (310 to 330).

[0095] Figure 8 An enlarged view of the daily contribution display area 330 for the selected date is shown. In the daily contribution display area 330, the left chart shows the contribution of each tag relative to the selected specific date in the order of tags that have the greatest impact on the tag value, with the vertical axis representing the tag of the main key factor and the horizontal axis representing the contribution of the tag. Figure 8 The right side of the graph shows the contribution of each tag in a cumulative line graph. Figure 8 From the chart, we can see that among the 10 key factors, as a single factor, label (T) has a significant effect on label ( Figure 8 The decline in conversion rate (expressed as a percentage of conversion rate) had the greatest impact.

[0096] In addition, in order to immediately confirm whether the analysis of such XAI is accurate and reliable, the system of the present invention includes a function of displaying actual detection data. Figure 7 As shown, each tag contribution display area 320 may include a "cause detailed analysis" selection key 10 and an "actual data" selection key 20. Figure 7 That is, in one embodiment, the "cause detailed analysis" screen can be used as a basic default output. At this time, if the user wants to confirm the actual data, he can click the "actual data (Actual Data)" selection key 20 to confirm.

[0097] like Figure 4 In step (S250), when the user clicks the "Actual Data" selection button 20, the user system outputs a screen where the user can set and confirm the target period, label, and target, and the user can select the label and target on this screen. For example, this screen can be further displayed on the current screen, or it can be displayed as a pop-up window. When the user selects a period, one or more labels and targets and sets them in step (S250), the actual observation data of the period is output in step (S260). For example, the system outputs Fig. 9 The data shown in Figure 1, where each chart shows the label and the observed value according to the label time. Fig. 9 As shown, the black graph (Gc) represents the target (conversion rate) value, and the orange graph (G T )express Figures 6 to 8 The observation value of the label (T) with the highest contribution in . Fig. 9 It can be confirmed that the actual enhancement trends of tags (T) and conversion rates are similar.

[0098] As described above, the present invention uses XAI to extract key factors, display the contribution of each key factor, and further analyze the influence of each key factor in detail. For example, Figure 7 In the tag contribution display area 320, when the target value of a specific date is extremely low, if the user selects the date on the UI screen (for example, Figure 7 The date indicated by arrow 30 in the output Figure 8 The daily contribution display area 330 allows the user to confirm the factors that currently affect the target change in order of contribution, and to analyze which factors cause the target value to drop. Figure 7 When you select the "Actual Data" option key 20 on the screen, the system directly outputs Fig. 9 The actual observation data can be used to verify the key factor contribution results under XAI analysis, and the reliability of XAI analysis can also be improved.

[0099] Again, reference Figure 2 It can be seen that, as described above, the key factors calculated by the root cause analysis of step (S20) can be applied to the subsequent yield prediction step (S30). For example, when more than 100 key factors are selected from more than 9,000 tags in the factory in step (S30), the subsequent yield prediction step (S30) can implement yield prediction based on the tag data values ​​of the selected more than 100 key factors.

[0100] A yield prediction model using a known machine learning method can be learned for yield prediction. In this case, the yield prediction model can be an existing statistical model or a prediction model based on machine learning or deep learning.

[0101] In one embodiment, the yield prediction model of the present invention can adopt a time series prediction model. For example, the time series prediction model can adopt traditional sequence models such as exponential smoothing model (ETS:Exponential Smoothing), autoregressive moving average model (ARIMA, AutoRegressive Integrated Moving Average); support vector machine (SVM:Support Vector Machine), random forest (Random Forest) XGBoost, LightGBM and other machine learning-based models; and recurrent neural network (RNN, Recurrent Neural Network), long short-term memory network (LSTM, Long Short Term Memory), GRU and other deep learning-based models One or more. For example, preferably, a bagging (Bagging) series algorithm such as random forest and a boosting (Boosting) series algorithm such as XGBoost, LightGBM can be integrated to adopt a yield prediction model with high predictive power.

[0102] The bagging algorithm is a method of performing random sampling and statistics in parallel for multiple iterations, which has the characteristic of increasing learning data. Therefore, even if the learning data is insufficient, sufficient learning effect can be obtained, which has the effect of preventing underfitting and overfitting.

[0103] The boosting algorithm also implements multiple random samplings, but it does not implement them in parallel, but in sequence. Therefore, it has the characteristic of adjusting the weight value of the next learning data based on the previous learning effect to achieve learning. That is, it gives a high weight value to wrong answers, so it has a high accuracy effect.

[0104] In the above-mentioned model, preferably, the bagging series algorithm for preventing overfitting and underfitting is learned, the shorthand forest algorithm or error is widely used, and the LGBM model in the boosting series algorithm for improving accuracy is used for modeling. However, this yield prediction model is schematic, and of course, a known suitable machine learning method can be used according to the specific embodiment of the present invention.

[0105] Alternative embodiments of the present invention are now described.

[0106] like Figure 2 As shown, in an alternative embodiment of the present invention, the yield prediction method does not directly use the pre-processed data for root cause analysis and yield prediction, but further implements at least one of the steps of segment analysis (S40), data realization (S50) and aging factor analysis (S60), and then applies it to root cause analysis and yield prediction. Figures 10 to 16 is a schematic diagram of a method for the segment analysis step (S40) according to an embodiment, Fig.10 is a schematic flow chart of an illustrative method of the segment analysis step, Figures 11 to 16 Schematic diagram of a segment analysis process according to an embodiment.

[0107] The segment analysis step (S40) divides a cycle into multiple segments based on the preprocessed data. In one embodiment, the segments can be divided into segments based on the inflection point where the yield trend changes sharply and the degree of yield change. The segments are divided into segments showing similar yield increase and decrease trends during the process operation period of a cycle, and different segments are applied in various modeling such as key factor extraction, yield prediction, and yield prediction simulation in the future, thereby improving the accuracy of yield prediction.

[0108] like Fig.10 As shown, the segment analysis step (S40) described in one embodiment may further include: a step of selecting the main factors required for segment analysis (S410), a step of selecting the inflection point of the main factors, and a step of determining the segment once (S420). In addition, in one embodiment, after step (S420), the segment is integrated or separated by variability analysis, and a step of determining the segment twice (S430). In addition, in one embodiment, after step (S430), a step of determining the segment three times based on catalyst design (S440) may be further included.

[0109] The steps are briefly described as follows: First, in step (S410), the main factors required for the segment analysis are selected. For example, the main factors can be selected, and the main factors include: at least one of the yield, conversion rate, and selectivity as target values, and the main tags that affect these target values. For this, Fig.11 The eight main factors selected in step (S410) are schematically shown.

[0110] Next, in the step of splitting into segments (S40), the inflection points of the main factors are searched and analyzed to determine the segments at one time. The specific method of analyzing the inflection points to search for the segments is as follows: (i) calculating the inflection points of one or more factors required for selecting the segments within a period, (ii) clustering the calculated inflection points, and (iii) selecting the inflection points that serve as segment boundaries from the clustered inflection points to determine the segments.

[0111] More specifically, in step (i), for example, the inflection point can be calculated using a known method such as a plateau detection method. Fig.12 The results of detecting the inflection point using the deformation stationary detection method are shown. Figure 6 The graph is a graph of the air temperature (Regen Air temperature) during catalyst regeneration. At this time, the X-axis represents the time within a cycle (or a variable corresponding to the time such as "cumulative production"), and the Y-axis represents the temperature.

[0112] Then, in step (ii), clustering is performed on the inflection points. For example, Fig.13 The density-based clustering algorithm (DBSCAN, Density-Based Spatial Clustering of Applications with Noise) that uses data density (density) to achieve clustering is shown, and the clustering achieved for the inflection point is shown. DBSCAN is one of the well-known clustering methods, and of course, the present invention is not limited to this method.

[0113] Next, among the clustered inflection points in step (iii), the inflection points that serve as segment boundaries are selected to determine the segments at once. For example, Figure 8 The result of determining the first segment in which the step (iii) is split into four segments is shown.

[0114] In one embodiment, the segment splitting result determined last time can be used to proceed to the next step (for example, Figure 2In the embodiment of the alternative scheme, after step (S420), the step (S430) of integrating or splitting the segments and re-determining the segments by variability analysis may be further included. In step (S430), for example, the average and deviation of different segments may be calculated based on the different segments determined once, and the segments determined once may be integrated or split based on the calculated average and deviation, and the segments may be determined twice. Fig. 9 The segment splitting result determined twice in step (S430) is shown. Fig.15 In FIG. 4 , the dotted line schematically represents the first segment splitting result obtained in step ( S420 ), and the blue solid line schematically represents the second segment splitting result obtained in step ( S430 ).

[0115] In one embodiment, the segment splitting result determined twice as above can be used to proceed to the next step (for example, Figure 2 Alternatively, after step (S430), a step (S440) of determining the sections three times based on the catalyst design may be further included. In step (S440), for example, the number of catalysts put into the process (e.g., 600,000 tons, 1.8 million tons, 2.4 million tons, etc.) is considered, and the similarity with the catalyst design is compared to finally determine the section splitting.

[0116] For example, Fig.16 The segment splitting result finally determined by this step (S440) is shown. Fig.16 As shown, when the catalyst life is regarded as one cycle (for example, 4 years), one cycle is divided into 4 sections (SG1 to SG4). Fig.16 In the figure, the red graph shows the air temperature during catalyst regeneration (Regen Air temperature), and the gray graph and the light green graph show the yield and selectivity, respectively.

[0117] like Fig.16 As shown, the first section (SG1) is a period in which a catalyst is added to gradually increase the Regen Air temperature and tend to stabilize. The second section (SG2) is a steady-state period in which the Regen Air temperature is maintained at a steady state, and the yield and selectivity are steady. From the third section (SG3), it can be seen that the yield gradually decreases with the aging of the catalyst. That is, although this step increases the Regen Air temperature, the yield will not be maintained or increased, but will gradually decrease due to the aging of the catalyst. The fourth section (SG4) is a step in which the yield further decreases sharply. Even if the Regen Air temperature is further increased, the yield and selectivity will not increase further, but will decrease sharply.

[0118] As described above, the present invention can consider the trend of increase or decrease in yield caused by catalyst life during a period, and can divide the intervals showing similar trends into segments, and according to the divided segments, in the subsequent steps (for example, Figure 2 The root cause analysis step (S20) and the yield prediction step (S30) shown in the figure) are respectively modeled according to different sections, and the overall analysis / prediction results are derived for one cycle, thereby improving the accuracy of the yield prediction.

[0119] Reference now Fig.17 and Fig.18 Describe the data realization steps ( Figure 2 The schematic method of S50 is shown in the figure. The data realization step (S50) generates and realizes the data for the remaining period of the second period based on the data of the first period and the first half of the data of the second period. The yield prediction step (S30) described in the present invention applies the data of the previous period (the first period) to predict the next period (the second period). More specifically, after the data of the second period is generated based on the data of the first period, the data of the second period is input into the yield prediction model to predict the yield of the second period. At this time, if there is data for a part of the second period (hereinafter, "the first half of the second period"), the data realization step (S50) applies the data of the first half of the second period to generate the data for the remaining period of the second period.

[0120] For example, the following is taken as a premise: a cycle period is 4 years, the first cycle is from January 2017 to December 2020, and the second cycle is from January 2021 to December 2024. Assuming that the current time point is August 2022, the data storage unit 200 stores the process operation data of the first cycle and the first half of the second cycle (i.e., from January 2021 to July 2022).

[0121] At this time, the data realization processing unit 150 generates data for the remaining period of the second cycle (i.e., from August 2022 to December 2024) based on the process operation data of the first cycle and the first half data of the second cycle. At this time, considering the trend or average characteristics of the first cycle data and the characteristics of the first half data of the second cycle, the data realization processing unit 150 can generate data for the remaining period of the second cycle.

[0122] However, due to differences in catalyst addition amount, initial operating conditions, etc., the process operation data (e.g., each tag data) of the first cycle and the second cycle tend to have different trends or values ​​in the two cycles. Therefore, it is difficult to generate data for the second cycle using the operating condition data of the first cycle. Therefore, the data realization step (S50) of the present invention corrects the data of the first cycle to make it consistent with the trend of the second cycle, thereby generating data for the second cycle.

[0123] In one embodiment, the method for realizing each tag data may at least include an average difference reflection method and a random number generation method.

[0124] When there is an average difference between the first cycle and the second cycle, the average difference reflection method can be applied. In one embodiment, when the data change trends in a given period of the first cycle and the second cycle are similar, but the average values ​​are different, the average point of the first cycle data is moved to generate the data of the second cycle. For example, Fig.17 Schematic label data to which the mean difference reflection method can be applied are shown. Fig.17 In the graph, the X-axis is the axis corresponding to the time of one cycle, and the Y-axis represents the data value of the tag. In addition, the black chart is the data of the first cycle of the tag (CY1), and the red chart is the first half of the data of the second cycle (CY21). It should be understood that the time point when the first half of the data of the second cycle (CY2) ends is the current time point.

[0125] Comparing the data of the first cycle (CY1) with the data of the first half of the second cycle (CY21), it can be seen that the trends of the two data are similar, and the average value of the data of the second cycle is larger. Therefore, at this time, the average difference reflection method can be applied to increase the data of the corresponding period of the first cycle to the average difference, thereby generating the data of the remaining period of the second cycle. At this time, in one embodiment, the data of the corresponding period of the first cycle is increased to the average and used directly. Alternatively, at least for a part of the interval, the data can also be modified by generating random numbers and other methods to generate the second data.

[0126] When the data for a given period is incomplete or contains outliers, a random number generation method can be applied. In one embodiment, when there is incomplete data in the first period, a random number is generated to generate data for the second period. For example, Fig.18 Schematic label data to which the random number generation method can be applied is shown.

[0127] Fig.18 In the chart, the X-axis is the axis corresponding to the time of one cycle, and the Y-axis shows the data value of the tag. The black chart is the data of the first cycle of the tag (CY1), and the orange chart is the first half of the data of the second cycle (CY21). The time point at which the first half of the data of the second cycle (CY2) ends means the current time point.

[0128] like Fig.18As shown, the first cycle data (CY1) is generated after a certain time point. That is, the tag data may mean that the sensor was not set or the sensor was not working before the specific time point. However, the first half of the second cycle data (CY21) exists, and the data of the remaining interval (CY22) is based on the data of the first cycle, and the second half of the data (CY22) is generated by generating random numbers. At this time, for example, the average and dispersion of the previous given interval (for example, the previous 30 days (D30) at the current time point) are calculated at the current time point, and while maintaining the average and dispersion, the second half of the data (CY22) is generated by generating random numbers for the data at individual moments.

[0129] Reference now Figures 19 to 21 Describe the steps of aging factor analysis ( Figure 2 The schematic method of step S60 is shown in FIG. 1 . The purpose of implementing the aging factor analysis step (S60) is to reflect the aging of the process catalyst, and the purpose of using the catalyst in the process is to more accurately predict the process yield. In the PDH process for producing propylene, as the cumulative production of propylene increases, the life of the catalyst is shortened, resulting in a sharp drop in the yield of the catalyst in the second half of the process. Therefore, in order to accurately reflect the yield change caused by the change in process conditions and the yield reduction caused by the life of the catalyst in the yield prediction, preferably, the aging factor of the catalyst is used to implement the yield prediction.

[0130] In one embodiment, as shown in the following formula, the aging factor of the catalyst converts the value obtained by dividing the daily propylene production by the heat applied to the catalyst into an index, reflecting the yield reduction value.

[0131] AF (Aging Factor) = (Daily Propylene Production) / (Heat Applied to Catalyst)

[0132] In the above formula, for example, the "heat amount applied to the catalyst" can be obtained by multiplying the tag data value indicating the Regen Air temperature by the flow rate.

[0133] For example, Fig.19 As shown, the yellow graph is the yield (Y1) of the first cycle in a cycle along with time, and the red graph represents the aging factor (AF1) of the first cycle. It can be seen that the aging factor (AF1) increases and decreases in a trend similar to the yield (Y1). Fig.16 The description made is that the fourth section is the last section of the cycle. No matter how the temperature is increased, the yield will decrease. At this time, until the third section, the yield decreases relatively linearly, while in the fourth section, the yield decreases nonlinearly.

[0134] in addition, Fig.19The green chart in the middle is the yield (Y2) of the second cycle up to the current time point, and the blue chart represents the aging factor (AF2) of the second cycle up to the current time point, extending in a trend similar to the yield (Y2) and aging factor (AF1) of each first cycle. Therefore, it can be predicted that the corresponding yield (Y2) during the fourth segment will still decrease nonlinearly. In order to make a more accurate prediction, the yield is calculated by reflecting the aging factor (AF) corresponding to the second cycle.

[0135] For example, Fig. 20 As shown, regarding the aging factor (AF1) of the first cycle, first, an average value (AF1m) can be calculated according to different sections, and the calculated average values ​​of each section can be applied to each section of the second cycle as a weighted value. In one embodiment, even if the aging factor (AF) is not applied to the first and second sections, the yield prediction model can still accurately predict the yield. Therefore, the aging factor can be not applied to the first and second sections, while the aging factor (AF) can be applied to the third and fourth sections. In another embodiment, the yield of the third section decreases more linearly, and the yield prediction model can be accurately predicted to a certain extent. Therefore, the aging factor (AF) can be applied only to the fourth section.

[0136] Fig.21 This is a diagram showing the trend of the predicted yield when the aging factor (AF) is not applied to the first and second segments and only the aging factor (AF) is applied to the third and fourth segments. If the aging factor (AF) is not applied to the third and fourth segments, the yield prediction is as follows Fig.21 In the blue graph shown, the difference between the predicted yield (blue solid line) and the actual yield (black dashed line) is not large in the first and second segments, but the error shown in the third and fourth segments is large. However, as described in the present invention, when the aging factor (AF) is applied to the third and fourth segments, as shown in the red dashed line, the yield prediction becomes closer to the actual yield.

[0137] In addition, if Figure 2 As shown, the implementation result of the aging factor analysis step (S60) can be applied in the implementation of the yield prediction simulation (80). Fig.16 The predicted yield (Y) resulting from the implementation result of the yield prediction step (S30) relative to the second cycle, as described above, represents the following technical content: that is, dividing a cycle into 4 sections, realizing data realization according to each different section, reflecting the aging factors of the 3rd and 4th sections, and implementing the yield prediction result.

[0138] As described above, the catalyst activity prediction system and method for reflecting key factors in a chemical process described in the present invention can screen key factors according to process operating conditions, apply them to the prediction model, and improve the prediction accuracy of catalyst activity changes.

[0139] Through the above process, the catalyst activity can be accurately predicted, the product sales plan and the timing of catalyst replacement can be determined, the time and cost required for product production and catalyst replacement can be effectively managed, the catalyst activity changes can be grasped according to the process operating conditions, the operating conditions can be optimized, the catalyst activity can be improved, and the product output can be increased.

[0140] As described above, any person skilled in the art can make various modifications and variations from the descriptions of these specifications. Therefore, the scope of the present invention should not be limited to the described embodiments, but should be defined by the following claims and equivalents thereof.

[0141] [Possibility of industrial application]

[0142] The present invention analyzes the root causes of chemical processes, and more specifically, relates to a yield prediction system and method for screening out key factors that drive yield changes and applying them to a yield prediction model.

Claims

1. A root cause analysis and yield prediction method, characterized in that: In the root cause analysis and yield prediction method using the chemical process yield prediction system, include; A step of preprocessing the yield prediction data (S10); and The step of analyzing the root cause using the pre-treated yield prediction data (S20), The root cause analysis step (S20) includes a step of extracting process key factors (S21) and a step of analyzing the process impact of the extracted key factors (S22).

2. The root cause analysis and yield prediction method according to claim 1, characterized in that: The step of analyzing the process influence of key factors (S22) further includes the following steps: visualizing the influence of a selected given tag through a user interface (UI) according to user input.

3. The root cause analysis and yield prediction method according to claim 1, characterized in that: Further including: A step (S30) of learning a yield prediction model based on the preprocessed yield prediction data and the extracted key factors.

4. The root cause analysis and yield prediction method according to claim 1, characterized in that: The data preprocessing step (S10) includes: A step of preprocessing the yield prediction data in units of minutes (S110); A step of selecting an analysis target tag (S120); A step of extracting time and day unit data from the data of the selected label in the data pre-processed in minutes (S130); and A step of interpolating abnormal values ​​and missing values ​​for the daily unit data (S140).

5. The root cause analysis and yield prediction method according to claim 4, characterized in that: The yield prediction data includes: process operation data of olefin producers; laboratory data including LIMS data; plant event data including data related to the time when the plant is not operating normally; and historical yield, conversion rate and selectivity data related to olefin production.

6. The root cause analysis and yield prediction method according to claim 3, characterized in that: Further including: A yield prediction implementation step (S40) of implementing yield prediction using the yield prediction model learned in the yield prediction model learning step, and outputting the yield prediction result; A yield prediction implementation step (S40) of implementing yield prediction using the yield prediction model learned in the yield prediction model learning step, and outputting the yield prediction result; as well as The yield prediction results are used to evaluate the performance of the yield prediction model, and the performance evaluation and optimal model screening steps (S50) are performed to screen the best prediction model.

7. The root cause analysis and yield prediction method according to claim 4, characterized in that: The root cause analysis step (S20) applies the pre-processed yield prediction data to a describable artificial intelligence (XAI) algorithm to extract the process key factors.

8. A computer-readable recording medium having a computer program recorded thereon for implementing the root cause analysis and yield prediction method according to any one of claims 1 to 7.

Images