Method and system for process control

Abstract

This invention relates to a method for controlling a chemical process, wherein the chemical process is one or more of the preparation of methanol, hydrogen sulfide, methanethiol, hydrogen cyanide, acrolein, 3-methylthiopropional, 5-(2-methylthioethyl)-hydantoin, methionine, methionine salts, and methionine derivatives, comprising the steps of: a) providing a training set TS1, wherein the training set TS1 includes process values ​​PV1 and PV2 associated with each other, and / or laboratory values ​​LV1 and PV2 associated with each other; b) training a processing unit based on the training set TS1 of step a) to identify one or more measured process variables and at least one... The correlation model between process variables is used to develop a calibration function CF1 for the calibrated soft sensor based on the identified correlation model. c) Predict at least one operating parameter for the chemical process as an approximation of LV1 and / or PV1. d) Calculate the deviation as the difference between the predicted operating parameter and the corresponding laboratory value LV1 and / or treatment value PV1 of the training set TS1 in step a). e) If the deviation calculated in step d) exceeds a threshold, proceed to step f); otherwise, proceed to step g). f) Recalibrate the soft sensor in step b). g) Write the predicted at least one operating parameter into the DCS. h) Repeat steps c) to g).

Description

[0001] This invention relates to the field of process control, particularly to the control of chemical processes, specifically one or more of the preparation of methanol, hydrogen sulfide, methanethiol, hydrogen cyanide, acrolein, 3-methylthiopropional, 5-(2-methylthioethyl)-hydantoin, methionine, methionine salts, and methionine derivatives. Process control is a combination of control engineering and chemical engineering disciplines that uses industrial control systems to achieve a level of product consistency, economy, and safety that is impossible through purely manual control. It is widely used in industries such as oil refining, pulp and paper manufacturing, chemical processing, and power plants. It ranges widely in scale, type, and complexity, but enables a small number of operators to manage complex processes with high consistency. The development of large-scale automated process control systems facilitates the design of high-volume and complex processes that would otherwise be uneconomical or unsafe to operate. Applications range from controlling the temperature and level of a single process vessel to entire chemical processing plants with thousands of control loops.

[0002] Process control in large industrial plants has evolved through many stages. Initially, control came from panels located locally within the processing plant. However, this required significant human resources to manage these dispersed panels and offered no overall view of the process. The next logical development was to transmit all plant measurements to a permanently manned central control room. This essentially centralized all localized panels, offering the advantages of lower staffing levels and easier process overview. Controllers were typically located behind control room panels, with all automatic and manual control outputs transmitted back to the plant. However, while providing a central control focus, this arrangement was inflexible because each control loop had its own controller hardware, requiring operators to constantly move within the control room to view different parts of the process.

[0003] With the advent of electronic processors and graphic displays, these discrete controllers could be replaced by computer-based algorithms hosted on input / output rack networks with their own control processors. These could be distributed around the plant and communicate with graphic displays in one or more control rooms. Thus, distributed control systems were born.

[0004] The introduction of a distributed control system (DCS) facilitates the interconnection and reconfiguration of plant controls (such as cascaded loops and interlocks) and their integration with other production computer systems. It enables sophisticated alarm handling, introduces automatic event logging, eliminates the need for physical logs (such as graphical recorders), allows control of racks to be networked, thus enabling localization within the plant to reduce cabling, and provides a high-level overview of plant status and production levels.

[0005] Furthermore, the introduction of DCS ensures the application of more sophisticated control methods that use mathematical optimization to calculate optimal operating parameters that match given constraints. However, it requires control models to govern the processing of the computer and DCS, particularly for predicting the behavior of the system under discussion. A control model is a set of equations used to predict system behavior and can help determine the response to changes. To determine the basic model of any process, the definitions of the system's inputs and outputs differ from those of other chemical processes. For example, equilibrium equations are defined by control inputs and outputs rather than material inputs. State variables (x) are measurable variables that can well indicate the system's state, such as temperature (energy balance), volume (mass balance), or concentration (component balance). Input variables (u) are specified variables, typically including flow rate.

[0006] Therefore, industrial processing plants are typically equipped with a large number of sensors. The primary purpose of sensors is to deliver data for process monitoring and control. Decades ago, researchers began to utilize the vast amounts of data measured and stored in process industries to build predictive models based on this data. However, it still requires measurement data to predict the behavior of the system in question. Another problem in this regard is that measuring certain parameters in processing plants is often quite difficult, i.e., time-consuming, complex, or even impossible. Here, so-called soft sensors provide support. A soft sensor, consisting of text software and a sensor, is a virtual sensor. Therefore, it is not a real sensor, but rather a simulation of the dependence of a representative single indicator or measurement on a target value. Thus, the target value is not directly measured, but calculated or approximated through measurements related to the target value and that correlation model. At a very general level, two different categories of soft sensors can be distinguished: model-driven soft sensors and data-driven soft sensors. Typically, model-driven soft sensors describe the physical and chemical background of the process. However, a real drawback of these models is that they are primarily developed for the planning and design of processing plants, and therefore usually focus on describing the ideal steady state of the process. Therefore, they do not reflect real-world conditions, which are characterized not by an ideal process steady state, but by constantly changing, and especially unexpectedly changing, process states. Consequently, model-driven soft sensors require a considerable amount of engineering to respond to disturbances. This leads to significant complexity and sometimes makes it impossible to account for all disturbances, especially when they are generally unknown and the causes of the disturbances are unknown. This limits the use of model-driven soft sensors for controlling chemical plants. In contrast, data-driven soft sensors are based on data measured within the processing plant, and therefore they describe real process conditions more realistically than model-driven soft sensors.

[0007] Nevertheless, regular maintenance and fine-tuning of soft sensors, even data-driven ones, remains necessary. Maintenance is essential because data drift and other variations can degrade soft sensor performance. Therefore, soft sensors must be compensated for by adjusting or redeveloping the model. However, current soft sensors do not offer any automated mechanisms for maintenance. Thus, manual control and maintenance of soft sensors are still required. As if that weren't bad enough, there is still no acceptable absolute metric to assess the quality level of soft sensors. Therefore, the judgment of whether a model is performing well still depends on the model operator's subjective perception based on a visual interpretation of the deviation between the correction target value and its prediction. However, this requires extensive experience on the control side. Worse still, this does not prevent the control operator from making incorrect decisions, as their judgment is based solely on their subjective perception rather than on an objective assessment of a higher level of situation. Things become even more complex when it comes to chemical processes representing the multi-step preparation of chemicals. Such multi-step processes are, for example, the preparation of methionine, which involves the preparation of several starting compounds and intermediate compounds.

[0008] Therefore, there is a need for a method that allows for the automatic control of chemical processes via soft sensors, wherein the chemical processes are one or more of the preparation of methanol, hydrogen sulfide, methanethiol, hydrogen cyanide, acrolein, 3-methylthiopropional, 5-(2-methylthioethyl)-hydantoin, methionine, methionine salts, and methionine derivatives, wherein the soft sensors are automatically recalibrated.

[0009] The problem has been resolved by applying automatically recalibrated soft sensors within the framework for controlling chemical processes. Specifically, when operating parameters (predicted for one or more process values ​​and associated with the process and / or laboratory values ​​as laboratory values ​​or approximations of process values) deviate from their corresponding laboratory and / or process values, an automatically recalibrated soft sensor is performed. This calibrated soft sensor is obtained by training a processing unit based on a training set of process and laboratory values ​​correlated with each other, and / or a set of process values ​​correlated with another set of process values. Specifically, in this case, the original training set used to provide the calibrated soft sensor is expanded with additional laboratory and / or process values, and the processing unit is trained based on this expanded training set to obtain a corrected calibration function with the aid of the recalibrated soft sensor. The resulting recalibrated soft sensor is then used to predict process parameters for controlling the chemical process. When there is no deviation between the predicted process parameters and their corresponding laboratory values, the predicted parameters for operating the chemical process are written into the distributed control system (DCS) of the process.

[0010] Therefore, one object of the present invention is a method for controlling a chemical process, wherein the chemical process is one or more of the preparation of methanol, hydrogen sulfide, methanethiol, hydrogen cyanide, acrolein, 3-methylthiopropional, 5-(2-methylthioethyl)-hydantoin, methionine, methionine salts, and methionine derivatives, comprising the following steps: a) Provide a training set TS1, wherein the training set TS1 includes process values ​​PV1 and PV2 that are correlated with each other, and / or laboratory values ​​LV1 and PV2 that are correlated with each other. b) Based on the training set TS1 obtained in step a), train the processing unit to identify correlation models between one or more measured process variables and at least one process variable, and develop a calibration function CF1 for the calibrated soft sensor based on the identified correlation models. c) Predicting at least one operating parameter for the chemical process as an approximation of LV1 and / or PV1, comprising the following steps: c1) Request one or more process values ​​corresponding to PV2 from the distributed control system (DCS) of the chemical process, and c2) Using the calibrated soft sensor described in step b), predict the operating parameters with the values ​​described in step c1). d) The bias is calculated as the difference between the predicted operating parameters in step c2) and the corresponding laboratory value LV1 and / or process value PV1 of the training set TS1 in step a). e) If the deviation calculated in step d) exceeds the threshold, then proceed to step f); otherwise, proceed to step g). f) Recalibrate the soft sensor described in step b), which includes the following steps: f1) Expand the training set TS1 of step a) with another laboratory value LV1 and / or another process value PV1, or replace at least a portion of the training set TS1 with another laboratory value LV1 and / or another process value PV1 to provide training set TS2. f2) Train the processing unit of step b) based on the expanded training set TS1 or the training set TS2 of step f1) to correct the calibration function of step b). f3) Recalibrate the soft sensor from step b) using the modified calibration function described in step f2), and f4) Return to step c) using the recalibrated soft sensor described in step f3). g) Write the at least one operating parameter predicted in step c2) into the DCS, and h) Repeat steps c) to g).

[0011] The method according to the invention is also Figure 1 The explanation is as follows.

[0012] Continuous execution of the method according to the invention is not required. Instead, it is appropriate to take a break after the at least one operating parameter predicted in step c2) has been written into the DCS in step g). This also allows the various systems of the chemical process to reach a steady state. Preferably, the method according to the invention is executed periodically at intervals of one to ten minutes, for example every minute, every two minutes, every three minutes, every four minutes, every five minutes, every six minutes, every seven minutes, every eight minutes, every nine minutes, or every ten minutes. More preferably, the method is executed periodically at the same time intervals in at least the same phase of the chemical process in question. However, changing the time interval is meaningful when transitioning from one particular phase of the chemical process in question to another, as this may cause significant changes in the chemical process.

[0013] In the context of this invention, the term soft sensor is used as known to those skilled in the art and refers to a predictive model based on data measured and stored in process industries. The term is a combination of the words "software" (because the model is typically a computer program) and "sensor" (because the model conveys information similar to its hardware counterpart). In principle, a soft sensor is a mathematical function of any shape, depending on its structure and the training or learning algorithm used to create it.

[0014] According to the invention, at least one operating parameter of the chemical process is predicted to be close to LV1 and / or PV1 in step c) of the method. This allows in step d) a comparison in which the predicted value (i.e., the operating parameter predicted in step c) is compared with the expected value (i.e., the corresponding laboratory value LV1 and / or treatment value PV1 of the training set TS1 in step a), wherein the deviation is calculated as the difference between the predicted value (i.e., the operating parameter predicted in step c) and the expected value (i.e., the corresponding laboratory value LV1 and / or treatment value PV1).

[0015] Specifically, step c) involves requesting one or more process values ​​from the distributed control system (DCS) of the chemical process, wherein the one or more process values ​​correspond to PV2. Therefore, the process value requested in step c1) is also related to the laboratory value LV1 and / or the process value PV1. Thus, a soft sensor based on a calibration function CF1 generated from the correlation between the laboratory value LV1 and the process value PV2 and / or between the process values ​​PV1 and PV2 is suitable for predicting at least one operating parameter as an approximation of LV1 and / or PV1.

[0016] In principle, the method according to the present invention utilizes historical data to extract correlations between process variables. Specifically, a processing unit is trained with a training set of historical data (i.e., process variable values ​​that are correlated with each other) to identify a correlation model between one or more measured process variables and at least one process variable to be predicted or approximated. Next, an equation (i.e., a calibration function for the soft sensor) is developed from the identified correlation model between the one or more measured process variables and the at least one process variable to be predicted or approximated. To develop the equation, i.e., the calibration function, a suitable training set is required. Specifically, in step a), a training set T1 is provided, comprising process values ​​PV1 and PV2 that are correlated with each other, and / or laboratory values ​​LV1 and PV2 that are correlated with each other. The first step in constructing the training set is to collect values ​​that are correlated with each other, i.e., laboratory value LV1 and process value PV2, process value PV1 and process value PV2, or both pairs.

[0017] To appropriately reflect the progress and changes of the process under discussion, it is beneficial to perform appropriate time-series assignment on the values ​​that are related to each other. Therefore, for each value—that is, each laboratory value LV1, each process value PV1, and each process value PV2—a time stencil is provided, indicating the time point at which the value was recorded and / or the sample for which the value was obtained. Establishing a time series of these values ​​is also beneficial. For this purpose, a time pattern is provided for each laboratory value LV1 and / or each process value PV1, tracing back a predetermined time span from the time stencil of LV1 and / or PV1. This invention is not limited by any specific time span. Rather, the predetermined time span is selected taking into account the corresponding periodic variations of the related values. For example, depending on the corresponding circumstances—the distance between the point where excess acid was added and the point where the pH value must be known, and the distribution of the acid in the linked system—values ​​of excess acid with pH influence are collected every 15, 30, 45, or 60 minutes. Finally, the related values ​​are linked when the time stencil of a value matches the time pattern of the corresponding related value. Specifically, when the time template of process value PV2 matches the time model of laboratory value LV1 and / or process value PV1, process value PV2 is linked with laboratory value LV1 and / or process value PV1. For example, process values ​​with time templates such as 12:15, 12:30, 12:45, etc., match laboratory value LV1 with a time model of 15 minutes, so they are linked.

[0018] In one embodiment of the method according to the invention, steps a) and / or f1) further include the following steps: A1) Collect laboratory value LV1 and process value PV2, and / or process value PV1 and process value PV2. A2) Provide a time template for each value in step A1), indicating the time point at which the value was recorded and / or the sample for which the value was obtained. A3) Provide a time model for each laboratory value LV1 and / or each process value PV1, and backtrack a predetermined time span from the time templates of LV1 and / or PV1. A4) Link laboratory value LV1 and / or process value PV1 to one or more process values ​​PV2 that have time templates that match the time model of laboratory value LV1 and / or the time model of process value PV1.

[0019] In the context of this invention, two values ​​are correlated when one value depends on or is influenced by another. An example of a process value depending on or being influenced by a laboratory value is one that depends on the pH of an excess acid. An example of a process value influencing a laboratory value is one that depends on the acid dosage of the excess acid. In the context of this invention, a laboratory value LV1 and / or a process value PV1 can be selected and predicted or approximated to obtain a process value PV2. Next, a list of labels is created that a) depend on or are influenced by a laboratory value, e.g., pH depends on an excess acid, or b) will influence a laboratory value, e.g., acid dosage affects an excess acid. According to this process, laboratory values ​​LV1 and / or process values ​​PV1 associated with process value PV2 can be collected.

[0020] In another embodiment of the method according to the invention, i) when process value PV1 depends on or is affected by process value PV2, process value PV1 and process value PV2 are related to each other, and vice versa, and / or ii) when laboratory value LV1 depends on or is affected by process value PV2, laboratory value LV1 and process value PV2 are related to each other, and vice versa.

[0021] Specifically, process value PV2 is collected and allowed to be associated with laboratory value LV1 and / or process value PV1. Therefore, the laboratory value LV1 and process value PV2 associated with each other depend on the chemical process to be controlled. Similarly, the process values ​​PV1 and process value PV2 associated with each other depend on the chemical process to be controlled. For example, if laboratory value LV1 or process value PV1 is indeed related to acid excess, then process value PV2 will be the pH value. For example, if laboratory value LV1 or process value PV1 is indeed related to ion concentration, then process value PV2 will be the ionic conductivity of the corresponding medium, such as the conductivity of methionine in the corresponding medium. For example, if the laboratory value LV1 or process value PV1 is indeed related to the concentration of the specific organic compound to be prepared, such as methanol, methanethiol, acrolein, 3-methylthiopropional, methionine, or a methionine derivative, such as 2-hydroxy-4-(methylthio)butyric acid, which is a hydroxy analog of methionine, then the process value PV2 will be the intensity of the characteristic absorption band of the organic compound in question (in IR, NIR, UV, or Raman spectra). Alternatively, if the laboratory value LV1 or process value PV1 is indeed related to the concentration of the specific organic compound that is the starting compound for the compound to be prepared (e.g., propylene for acrolein, methanol or hydrogen sulfide for methanethiol, 5-(2-methylmercaptoethyl)-hydantoin for methionine), then the process value PV2 will be the intensity of the characteristic absorption band of the organic starting compound in question (in IR, NIR, UV, or Raman spectra). For example, if the laboratory value LV1 or the process value PV1 is the oxygen concentration, then the process value PV2 will be the oxygen value obtained from electrochemical measurements such as current measurements or resistance measurements, or from optical measurements such as absorption or fluorescence measurements.

[0022] Preferably, the laboratory value LV1 and / or process value PV1 is one or more of the group consisting of acid excess, ion concentration, concentration of the specific organic compound to be prepared (e.g., methanol, methanethiol, acrolein, 3-methylthiopropanal, methionine or methionine derivatives such as 2-hydroxy-4-(methylthio)butyric acid), and oxygen concentration, and the process value PV2 is pH value, ionic conductivity, intensity of the characteristic absorption band of the organic compound in question (in IR, NIR, UV or Raman spectra), and oxygen value obtained from electrochemical measurements such as current measurements or resistance measurements or from optical measurements such as absorption or fluorescence measurements.

[0023] The corresponding process value can be the current value, that is, the value obtained immediately from the measurement results during the operation, or it can be the average value, that is, the value collected at different time points over a period of time, and then averaged to give a single value.

[0024] In other embodiments of the method according to the invention, the process value is a current value or an average value.

[0025] The advantage of using an average process value is that the noise is lower than that of the current process value. Therefore, it is preferable that the process value in the method according to the invention is an average process value.

[0026] In a preferred embodiment of the method according to the invention, the average value is obtained by averaging the aggregated process values ​​over a predetermined time span.

[0027] It is advantageous when the average value also meets the requirement that the time template of the average process value PV2 matches the time model of the laboratory value LV1 and / or the process value PV1. Therefore, it is preferable to average the aggregated or collected process values ​​over a predetermined time span in step A4.

[0028] In an alternative preferred embodiment of the method according to the invention, the average value is obtained by averaging the aggregated process values ​​over a predetermined time span in step A4.

[0029] Furthermore, the training in step b) of the method according to the invention can also take into account the residence time of relevant components, such as the starting compound or the compound to be prepared as described above. The residence time can be measured or obtained from an equation, particularly as the quotient of the reactor volume or equipment volume and the discharge volume flow. This method allows for the consideration of time shifts in the chemical reaction.

[0030] In this case, the training process in step b) may include the following steps: Step 1: Create a training set TS for a specific time period T of historical data, with a sampling frequency of f_s (e.g., one sample every 1 to 10 minutes).

[0031] Step 1.1: Consider the dwell time during the process to introduce a time delay into the training set. Scenario A: Dwell time is not considered when creating the training set, so no time delay is introduced between process values ​​of different process labels.

[0032] Scenario B: Consider the dwell time, assuming that the dwell time is constant throughout the entire time.

[0033] Scenario C: The process value label list will be used to derive the residence time using the corresponding formula. Due to process dynamics and tank level fluctuations, the residence time is time-dependent.

[0034] Step 1.2: Set the time period T = T2 - T1, where T1 is the start time and T2 is the end time.

[0035] Scenario A: No measures were taken. Scenario B: No action was taken. Case C: Read the process value PV2 of the process value tag at T = T2 - T1 - TR, where TR is greater than the maximum assumed or given residence time in the process. This gives the set TPVR, where TPVR = PVR(TR) and the sampling points: start time,..., start time + TR according to the selected sampling frequency.

[0036] A list of process tags (PVL) related to the laboratory value LV obtained by heuristic compilation or based on statistical methods such as principal component analysis is provided, which is the target value of the soft sensor. The laboratory values are collected over time T. In most cases, the sampling frequency f_sl of the laboratory values is much smaller than f_s, for example, f_sl = ¼ hour. f_sl << f_s holds. The historical process values using the process tags are read in TR, generating the set TPVL.

[0037] Step 1.3: Introduce a time delay Case A: No action Case B: The time delay in TPVL is introduced according to the provided constant residence time.

[0038] Case C: For each time stamp, the residence time is derived using TPVR. The time delay in TPVL is introduced according to the derived residence time.

[0039] Step 1.4: Assign TPVL to the time stamp of the laboratory value LV. If for the time t_LVi of the laboratory value LVi and the time stamp t_TPVL of the TPVL sample, t_TPVL < t_LVi and |t_TPVL – t_LVi| < T_L holds, then assign the laboratory value LV to the TPVL sample, and the measured or derived time interval is T_L. In this way, the number of samples composed of (TPVL, LV) is extended from T * f_sl to T_L * T * f_sl.

[0040] Step 2: Use TS in step b) of the method.

[0041] Create different training sets for each of Cases A to C. Each of them ultimately results in a different soft sensor. Therefore, in the process shown above, a total of three different training sets or a total of three different soft sensors can be created.

[0042] The identification and handling of outliers is always a critical point. Outliers may be due to variability in measurement or may be the result of instrument errors; the latter are sometimes excluded from the dataset. Generally, especially in statistical data, outliers are considered to be data points that are significantly different from other observations, but this leaves a lot of room for subjective and misinterpretations. On the other hand, including data points at the edges of the dataset is necessary for meaningful and robust calibration, and therefore they should not be simply skipped over just because they look odd. It has been shown that the problem associated with outliers (if not the main problem) is their detection or identification, because there is no rigid definition of outliers. Therefore, ultimately, determining whether an observation is an outlier remains a subjective act. Due to the lack of a universally accepted definition of outliers, there are various methods for detecting outliers. In the context of this invention, whether it is a laboratory value LV1, a process value PV1, or a process value PV2, it is considered an outlier when the variance of the measured value from its expected value is greater than 2σ. When the variance of the measured value from the expected value is greater than 2σ, the measurement based on the measured value is repeated. In this case, the terms variance, 2σ, and expected value are used as known in stochastic. Specifically, the term 2σ comes from the normal distribution, also known as the Gaussian or Gaussian distribution, where the variance, represented by σ, describes the width of the standard deviation. Roughly speaking, 99.45% of all measurements fall within a range of + / - 2σ from the expected value.

[0043] In other embodiments of the method according to the invention, when the variance of the value in question from its expected value is greater than 2σ, the measurements of the laboratory value LV1, process value PV1, and / or process value PV2 are repeated.

[0044] Preferably, the threshold in step e) of the method according to the invention is given by a relative error. Given a certain value... and its approximate value The absolute error is , where the vertical line represents the absolute value. If The relative error is The percentage error is In other words, absolute error is the difference between the exact value and its approximation, relative error is the absolute error divided by the exact value, and percentage error is the relative error expressed as a percentage. In principle, the method according to the invention is not subject to any restrictions regarding a specific threshold. Instead, the choice of threshold depends on the estimated variance of the approximation. Not wishing to be limited by a particular theory, the threshold in the method according to the invention is preferably 10% of the relative error.

[0045] In addition to the automatic recalibration of the soft sensor in step f), the recalibration can also be triggered under specific event circumstances. For example, automatic recalibration may be requested periodically at predetermined time intervals, or irregularly during predetermined stages of a chemical process, or when changing from one stage to another. For example, predetermined stages of a chemical process may be the start-up phase, continuous or batch operation, and / or shutdown phase of a production process. One example of changing from one stage to another is changing from continuous operation to batch operation. Other examples of predetermined stages of a chemical process may be the various stages of a purification process, such as the stages of separating a mixture of substances based on the different boiling points of the components during distillation or based on the different migration rates of the components in a colloidal matrix.

[0046] In one embodiment of the method according to the invention, step f) is also performed periodically at predetermined intervals, or irregularly during predetermined stages of a chemical process, or when changing from one stage to another.

[0047] As a supplement to or alternative to the automatic recalibration of the soft sensor in step f), one may also request that the recalibration of the soft sensor be triggered manually.

[0048] In an alternative embodiment of the method according to the invention, step f is triggered whenever deemed necessary.

[0049] According to the present invention, the soft sensor is recalibrated by sequentially f1) expanding the training set TS1 of step a) with additional laboratory values ​​LV1 and / or additional process values ​​PV1.

[0050] During the recalibration of the soft sensor in step f), the training set TS1 of step a) is expanded with additional laboratory values ​​LV1 and / or additional process values ​​PV1 to provide a new training set TS2. Therefore, step f) also involves collecting additional laboratory values ​​LV1, process values ​​PV1, and process values ​​PV2 to place these values ​​in a time series and frame, and to link the corresponding values ​​that are associated with each other. All these sub-steps are performed in the same manner as described in steps A1) through A4) above. Therefore, the linking of the additional values ​​in step A4) also allows for the provision of a separate training set TS2, which at least partially or completely replaces the training set TS1.

[0051] In another embodiment of the method according to the invention, in step f), the training set TS1 is at least partially or completely replaced by the training set TS2.

[0052] Conditions in chemical processes undergo steady changes, particularly those at the initiation stage (e.g., at the start of a chemical reaction) which are typically significantly different from those in later stages (e.g., during stable chemical reactions or processes). Therefore, if training set TS1 is partially replaced with training set TS2 in step f), it is preferable to replace the portion of training set TS1 representing the oldest conditions of the corresponding chemical process with the most recent values ​​(i.e., the portion of training set TS1 with the oldest time template value).

[0053] In a preferred embodiment of the method according to the invention, the laboratory value LV1 and / or process value PV1 having the oldest time template are replaced with the laboratory value LV1 and / or process value PV1 having the current time template.

[0054] Preferably, in step f), the laboratory value LV1 and / or process value PV1 with the oldest time template are at least partially or completely replaced with the laboratory value LV1 and / or process value PV1 with the current time template.

[0055] Given that specific phases of a chemical process typically have significantly different conditions, it is reasonable to replace training set TS1 at least partially or completely with training set TS2 at predetermined phases of the chemical process (e.g., after the start-up phase, in the shutdown phase, etc.). Alternatively, it is also reasonable to replace the existing training set TS1 at least partially or completely with training set TS2 when step f) is performed more frequently than is considered acceptable, i.e., when the number of times step f) is performed exceeds a defined threshold within a predetermined time period. Generally, the method according to the invention is not subject to any limitations regarding the threshold in step f). Rather, the threshold can be properly selected by taking into account the framework conditions and requirements to be met for each chemical process.

[0056] In other embodiments of the method according to the invention, TS1 is at least partially or completely replaced by training set TS2 during a predetermined phase of the chemical process or when the number of times step f) is executed exceeds a defined threshold within a predetermined time period.

[0057] According to the present invention, processing units are trained based on a training set to provide a calibration function. Typically, the processing unit is an artificial neural network. In the context of this invention, the term neural network is used synonymously with the term artificial neural network (ANN) and refers to a computational system inspired by, but not equivalent to, the biological neural networks that constitute the animal brain. Specifically, an ANN is based on a set of connection units or nodes called artificial neurons, which roughly mimic neurons in a biological brain. Each connection, like a synapse in a biological brain, can transmit a signal to other neurons. Artificial neurons receive signals, process them, and can send signals to the neurons connected to them. The "signal" at a connection is a real number, and the output of each neuron is calculated by some nonlinear function of the sum of its inputs. Connections are called edges. Neurons and edges typically have weights that adjust as learning progresses. Weights increase or decrease the signal strength at the connection. Neurons may have a threshold so that a signal is only sent when the aggregated signal exceeds that threshold. Typically, neurons are clustered into layers. Different layers can perform different transformations on their inputs. Signals are transmitted from the first layer (input layer) to the last layer (output layer), possibly after multiple traversals of these layers. For example, the layer that receives external input, such as the process value PV2, is the input layer. The layer that produces the final result, i.e., the predicted operating parameters that approximate LV1 and / or PV1, is the output layer. Hidden layers lie between them. There may be various connection models between the two layers.

[0058] In one embodiment of the method according to the present invention, the processing unit is an artificial neural network.

[0059] Artificial neural networks can be deep neural networks, recurrent neural networks, or convolutional neural networks, depending on the specific framework conditions and requirements to be met for the corresponding chemical process.

[0060] In the context of this invention, the term deep neural network or its abbreviation DNN is used as known to those skilled in the art and refers to an artificial neural network with multiple layers between the input and output layers. Different types of neural networks exist, but they always consist of the same components: neurons, synapses, weights, biases, and functions. These components function similarly to the human brain and can be trained like any other machine learning algorithm.

[0061] However, like many artificial neural networks, coarsely trained DNNs can suffer from numerous problems. Two common issues are overfitting and computation time. Due to the added layers of abstraction, DNNs are prone to overfitting, which allows them to model rare dependencies in the training data. In statistics, overfitting is an analysis that corresponds too closely or precisely to a particular dataset and therefore may fail to fit other data or reliably predict future observations. An overfitted model is a statistical model that includes more parameters than the data can demonstrate. The result of overfitting is the unwitting extraction of residual variations, or noise, as if these variations represented the underlying model structure. In other words, the resulting model memorizes a large number of instances rather than learning attentional features. Furthermore, DNNs must consider many training parameters, such as size (number of layers and number of units per layer), learning rate, and initial weights. Due to the time and computational resource costs, sweeping through the parameter space to find the optimal parameters may be impractical.

[0062] Here, the use of Convolutional Neural Networks (CNNs) is helpful. The term Convolutional Neural Network, or its abbreviation CNN, is used as known to those skilled in the art and refers to a class of deep neural networks that use a mathematical operation called convolution in at least one layer instead of general matrix multiplication. CNNs are a regularized form of multilayer perceptrons. Multilayer perceptrons typically represent fully connected networks, where each neuron in one layer is connected to all neurons in the next layer. The “complete connectivity” of these networks makes them prone to overfitting data. Typical ways of regularization involve adding some form of size measure of the weights to the loss function. CNNs take a different approach to regularization: they utilize hierarchical models in the data and assemble more complex models using smaller and simpler models. Thus, CNNs are at the lower end in terms of connectivity and complexity. Therefore, CNNs are less prone to overfitting and require less computational resources and less training data compared to other artificial neural networks. In the context of this invention, the need for less training data is a significant advantage, as in chemical processes, samples are sometimes collected only once an hour, resulting in less data collected than in other disciplines. However, high-quality prediction models, calibration functions, and soft sensors can be obtained using convolutional neural networks.

[0063] Therefore, convolutional neural networks are the preferred type of artificial neural network.

[0064] In the context of this invention, the term Recurrent Neural Network (RNN) or its abbreviation RNN is used as known to those skilled in the art and refers to a class of artificial neural networks in which the connections between nodes form a directed graph along a time series. This allows it to exhibit temporal dynamic behavior. The term "recurrent neural network" is arbitrarily used to refer to two main classes of networks with similar general structures: finite-pulse and infinite-pulse. Both classes of networks exhibit temporal dynamic behavior. A finite-pulse recurrent network is a directed acyclic graph that can be unfolded and replaced with a strictly feedforward neural network, while an infinite-pulse recurrent network is a directed cyclic graph that cannot be unfolded. Both finite-pulse and infinite-pulse recurrent networks can have additional stored states, and the storage can be directly controlled by the neural network.

[0065] According to the invention, the processing unit is trained on a training set TS1 to generate a calibrated soft sensor based on a calibration function CF1. Within the framework of controlling a chemical process by means of the method according to the invention, the training set TS1 can be provided by a calibration branch (1) for generating the calibration function, including a laboratory management system (LIMS) (2) providing laboratory values ​​LV1 (3) and a process information management system (PIMS) (5) for providing process values ​​PV1 and / or PV2 (6). For example, the so-called laboratory information management system (LIMS) is polled to obtain laboratory values ​​derived from experiments run in a laboratory representing the process in question. The training set (8) is created by linking laboratory values ​​LV1 with corresponding process values ​​PV2. Alternatively, process values ​​PV1 associated with at least one process value PV2 can also be collected to create the training set by linking process values ​​PV1 with appropriate process values ​​PV2. Providing a time template for each collected value facilitates linking laboratory values ​​LV1 with appropriate process values ​​PV2 or linking process values ​​PV1 with appropriate process values ​​PV2. The time template indicates the time point at which the corresponding value was recorded and / or the sample to which the value was recorded. The training set (8) thus generated is then used to train the processing unit (13) to generate a calibration function that allows the processing unit to perform predictions of operating parameters as approximations of the laboratory value LV1 and / or the process value PV1.

[0066] According to the invention, at least one operating parameter is predicted as an approximation of a laboratory value LV1 and / or a process value PV1 of a chemical process, wherein one or more process values ​​correspond to a process value PV2. Within the framework of controlling a chemical process by means of the method according to the invention, the one or more process values ​​can be provided by an operating loop (9) that requests the process parameter PV2 (12) from a distribution control system (DCS) (10) of the chemical process.

[0067] A processing unit (13) suitable for performing the method according to the invention is placed between the calibration branch (1) and the operating loop (9). Thus, the processing unit (13) is connected to the calibration branch (1), which provides the processing unit with the necessary laboratory value LV1, as well as the processing value PV1 and / or PV2 to generate a calibration function, and the operating loop (9) provides the processing unit with the necessary process value PV2, through which a prediction of at least one operating parameter approximating LV1 and / or PV1 is made.

[0068] Another object of the method according to the invention is a system for controlling a chemical process, comprising: i) The calibration branch (1) for generating the calibration function includes a laboratory information management system (LIMS) (2) for providing laboratory values ​​LV1 (3) and a process information management system (PIMS) (5) for providing process values ​​PV1 and / or PV2 (6). ii) An operating loop (9) for requesting one or more process values ​​(12) from a distributed control system (DCS) (10) of a chemical process, and iii) A processing unit (13) adapted to perform the method according to the invention, wherein the processing unit is connected to the calibration branch (1) and the operation loop (9).

[0069] Preferably, the system according to the invention is used to control the preparation of methanol, hydrogen sulfide, methanethiol, hydrogen cyanide, acrolein, 3-methylthiopropional, methionine, methionine salts or methionine derivatives.

[0070] In one embodiment of the device according to the invention, the processing unit is an artificial neural network.

[0071] The artificial neural network can be a deep neural network, a recurrent neural network, or a convolutional neural network, depending on the specific framework conditions and requirements to be met for the corresponding chemical process. Preferably, the processing unit is a convolutional neural network.

[0072] Optionally, the calibration branch further includes a laboratory value buffer (4) to which the laboratory value (3) is transferred after it is recorded. When the laboratory value buffer is full, the corresponding process value PV2 (5) is collected from the process information management system (5) for each laboratory value by a collector (7) and a training set is generated by linking the laboratory value LV1 and / or the process value PV1 with the matching process value PV2.

[0073] Then, the predicted operating parameters (14) are written into the distributed control system (10) of the chemical process by means of open platform communication (11).

[0074] This invention utilizes Figures 1 to 4 Further explanation is provided in the examples and embodiments.

[0075] Attached Figure Figure 1 This is a flowchart of the method according to the present invention.

[0076] Figure 2 This is a schematic diagram of the method and apparatus according to the present invention, wherein each number has the following meanings: (1) calibration branch, (2) laboratory information management system (LIMS), (3) laboratory value, (4) laboratory value buffer, (5) process information management system (PIMS), (6) process value, (7) collector, (8) training set, (9) operating loop, (10) distributed control system (DCS), (11) open platform communication (OPC), (12) process value, (13) processing unit and (14) predicted operating parameters.

[0077] Figure 3 This is a result diagram of a comparative embodiment.

[0078] Figure 4 This is a result diagram based on Embodiment 1 of the present invention.

[0079] Figure 5 This is a result diagram based on Embodiment 2 of the present invention.

[0080] Figure 6 'a' is the relative error predicted by the soft sensor not according to the present invention.

[0081] Figure 6 b is the relative error predicted by the automatic recalibration soft sensor according to the present invention.

[0082] Comparative Example: In the ammonia scrubber downstream of the reactor used to produce hydrogen cyanide, the excess sulfuric acid was predicted using a soft sensor and was not recalibrated. Furthermore, the excess sulfuric acid was also measured simultaneously as the actual value. Figure 3 This is a graph showing the predicted and actual measurement results, where the excess sulfuric acid value is measured in the laboratory in one case (solid black line, laboratory) and approximated by a soft sensor using existing technology in another case (dashed black line, soft sensor). The first ellipsis (dashed black line, left) indicates peak formation, and the second ellipsis (dashed black line, right) indicates offset correction. Figure 3 As can be seen, the process values ​​predicted by the soft sensor are close to the actual measured values. After an initial significant deviation, the predicted values ​​follow the trend of the measured values, but they never match the measured values. Instead, after a period of synchronization, the predicted values ​​begin to differ more from the actual values, and due to the large difference between the predicted and actual values, offset correction must be performed.

[0083] According to Embodiment 1 of the present invention: The excess sulfuric acid in the ammonia scrubber downstream of the reactor used to produce hydrogen cyanide is predicted by means of the method according to the invention. Furthermore, the excess sulfuric acid is also measured as the actual value. Figure 4 This is a graph showing the predicted and actual measurement results, where the excess sulfuric acid value is measured in the laboratory in one case (solid black line, laboratory), and approximated in another case by the method according to the invention (dashed gray line, predicted). Three ellipses represent the difference between the approximate and actual values, as well as immediate corrections. Figure 4 As can be seen, the process values ​​predicted by the method according to the invention have much better consistency with the actual measured values. Furthermore, the method according to the invention can also identify the difference between the predicted and actual values ​​quite quickly and recalibrate the soft sensor, thereby quickly restoring very good consistency between the predicted and actual values ​​after the deviation is identified.

[0084] According to Embodiment 2 of the present invention: This embodiment illustrates the recalibration of a soft sensor in the method according to the invention. Similarly, the excess sulfuric acid in an ammonia scrubber downstream of the reactor used for hydrogen cyanide production was predicted using the method according to the invention and measured in the laboratory. However, compared to Example 1, hydrogen cyanide production was shut down and then restarted. After the restart, the calibration function of the soft sensor no longer matched the conditions of the process. Therefore, there was a significant deviation between the sulfuric acid value measured in the laboratory and the predicted value of sulfuric acid. Figure 5 This large shift was visible from February 1, 2020 to August 1, 2020. However, once automatic training was initiated, i.e., recalibrating the soft sensor (indicated by the dashed line), the prediction of sulfuric acid values ​​improved significantly. From August 1, 2020 onwards, no further shift was observed. In the rare instances where there were discrepancies between predicted and actual values, the soft sensor was automatically recalibrated again, and the predicted values ​​once again showed good agreement with the actual values. The results are shown in... Figure 5 In this context, the value of excess sulfuric acid is measured in a laboratory setting in one case (cross, laboratory), and approximated by a method according to the invention in another case (black solid line, prediction).

[0085] Figure 6 a and 6b show the process before initiating recalibration (i.e., before retraining). Figure 6 a) and the relative prediction error after initiating recalibration (i.e., after retraining). Figure 6 b). Figure 6The results show that, if trained properly, soft sensors can typically make predictions with an error rate of 5% to 35%. However, the relative error before retraining is quite high, for example, 25% relative prediction error after 12 predictions, but the relative error will never be zero. Furthermore, the relative error of a soft sensor without automatic recalibration appears somewhat chaotic, particularly its unbalanced distribution and lack of a Gaussian distribution.

[0086] In comparison, Figure 6 b indicates that, compared to unrecalibrated soft sensors, automatically recalibrated soft sensors result in improved relative errors after retraining (i.e., initiating recalibration). Specifically, the relative prediction error of automatically recalibrated soft sensors is between -10% and +10%, thus the absolute value is significantly reduced. A major improvement is that the relative error is 0 for the majority of predictions. Furthermore, compared to... Figure 6 The error distribution in a forms a contrast. Figure 6 The error distribution in b is well balanced and conforms to a Gaussian distribution.

Claims

1. A method for controlling a chemical process, wherein the chemical process is one or more of the preparation of methanol, hydrogen sulfide, methanethiol, hydrogen cyanide, acrolein, 3-methylthiopropional, 5-(2-methylthioethyl)-hydantoin, methionine, methionine salts, and methionine derivatives. Includes the following steps a) Provide a training set TS1, wherein the training set TS1 includes laboratory values ​​LV1 and process values ​​PV2 that are correlated with each other. b) Based on the training set TS1 obtained in step a), train the processing unit to identify correlation models between multiple measurement process variables and at least one process variable, and develop a calibration function CF1 for the calibrated soft sensor based on the identified correlation models. c) Predict at least one operating parameter for the chemical process as an approximation of LV1, Includes the following steps c1) Request one or more process values ​​corresponding to PV2 from the distributed control system (DCS) of the chemical process, and c2) Using the calibrated soft sensor described in step b), predict the operating parameters with the one or more process values ​​described in step c1). d) The bias is calculated as the difference between the predicted operating parameter in step c2) and the corresponding laboratory value LV1 of the training set TS1 in step a). e) If the deviation calculated in step d) exceeds the threshold, then proceed to step f); otherwise, proceed to step g). f) Recalibrate the soft sensor described in step b), Includes the following steps f1) Expand the training set TS1 of step a) with additional interrelated laboratory values ​​LV1 and process values ​​PV2, or replace at least a portion of the training set TS1 with additional interrelated laboratory values ​​LV1 and process values ​​PV2 to provide training set TS2. f2) Train the processing unit of step b) based on the expanded training set TS1 or said training set TS2 from step f1) to correct the calibration function of step b). f3) Recalibrate the soft sensor from step b) using the modified calibration function from step f2), and f4) Return to step c) using the recalibrated soft sensor from step f3). g) Write the at least one operating parameter predicted in step c2) into the DCS, and h) Repeat steps c) to g).

2. The method according to claim 1, wherein step a) and / or f1) further comprises the following steps: A1) Collect laboratory value LV1 and process value PV2. A2) Provide a time template for each value in step A1), indicating the time point at which the value was recorded and the time point at which the value was obtained from the sample. A3) Provide a time model for each laboratory value LV1, and trace back a predetermined time span from the time template of LV1. A4) Link laboratory value LV1 with one or more process values ​​PV2 that have a time template that matches the time model of laboratory value LV1.

3. The method according to claim 1 or 2, wherein i) when the laboratory value LV1 depends on or is affected by the process value PV2, the laboratory value LV1 and the process value PV2 are correlated with each other, and vice versa.

4. The method according to claim 1 or 2, wherein the laboratory value LV1 is one or more of acid excess, ion concentration, concentration of the specific organic compound to be prepared, and oxygen concentration, and the process value PV2 is pH value, ionic conductivity, intensity of the characteristic absorption band of the organic compound under discussion, and oxygen value obtained from electrochemical measurement or from optical measurement; wherein the characteristic absorption band is a characteristic absorption band in IR, NIR, UV, or Raman spectroscopy.

5. The method according to claim 1 or 2, wherein the process value is a current value or an average value.

6. The method of claim 5, wherein the average value is obtained by averaging the aggregated process values ​​over a predetermined time span.

7. The method of claim 5, wherein the average value is obtained by averaging the aggregated process values ​​over a predetermined time span in step A4.

8. The method of claim 4, wherein the training in step b) further takes into account the residence time of the starting compound of the particular organic compound to be prepared.

9. The method of claim 4, wherein the training in step b) further takes into account the residence time of the specific organic compound to be prepared.

10. The method according to claim 1 or 2, wherein the measurements of laboratory value LV1 and process value PV2 are repeated when the variance between the laboratory value LV1 and / or process value PV2 in question and their expected value is greater than 2σ.

11. The method according to claim 1 or 2, wherein the threshold in step e) is given by a relative error.

12. The method according to claim 1 or 2, wherein step f) is also performed periodically at predetermined intervals, or irregularly during predetermined phases of the chemical process, or when changing from one phase to another.

13. The method of claim 2, wherein the laboratory value LV1 having the oldest time template is replaced with the laboratory value LV1 having the current time template.

14. The method of claim 2, wherein in step f), the laboratory value LV1 having the oldest time template is partially or completely replaced with the laboratory value LV1 having the current time template.

15. The method according to claim 1 or 2, wherein when the number of times step f) is executed exceeds a defined threshold within a predetermined time period, the training set TS1 is partially or completely replaced by the training set TS2 in a predetermined phase of the chemical process.

16. The method according to claim 1 or 2, wherein the processing unit is an artificial neural network.

17. The method of claim 16, wherein the artificial neural network is a deep neural network, a recurrent neural network, or a convolutional neural network.

18. A system for controlling a chemical process, comprising: i) The calibration branch (1) for generating the calibration function includes a laboratory information management system (LIMS) (2) for providing laboratory values ​​LV1 (3) and a process information management system (PIMS) (5) for providing process values ​​PV2 (6). ii) an operating loop (9) for requesting one or more process values ​​(12) from the distributed control system (DCS) (10) of the chemical process, and iii) A processing unit (13) adapted to perform the method according to claim 1 or 2, wherein the processing unit is connected to the calibration branch (1) and the operation loop (9).

19. The system of claim 18, wherein the system is used to control the preparation of methanol, hydrogen sulfide, methanethiol, hydrogen cyanide, acrolein, 3-methylthiopropional, methionine, methionine salt or methionine derivative.

20. The system of claim 18, wherein the processing unit is an artificial neural network.

21. The system of claim 20, wherein the artificial neural network is a deep neural network, a recurrent neural network, or a convolutional neural network.

22. The system of claim 18, wherein the calibration branch further comprises a laboratory value buffer (4) to which the laboratory value is transferred after the laboratory value (3) is recorded.

23. The system according to claim 18, wherein when the laboratory value buffer is full, a corresponding process value PV2 (6) is collected from the process information management system (5) for each laboratory value by a collector (7) and a training set is generated by linking the laboratory value LV1 with the matching process value PV2.

24. The system of claim 18, wherein the predicted operating parameters (14) are written into the distributed control system (10) of the chemical process by means of open platform communication (11).

Images