Process control method and process control system

Abstract

The present invention relates to a method for controlling a chemical process, wherein the chemical process is one or more of methanol, hydrogen sulfide, methyl mercaptan, hydrocyanic acid, acrolein, 3-methylthiopropionaldehyde, 5-(2-methylmercaptoethyl)-hydantoin, methionine, salts of methionine, and derivatives of methionine, the method comprising: a) providing a training set TS1, wherein said training set TS1 comprises mutually correlated process values ​​PV1 and PV2 and / or mutually correlated laboratory values ​​LV1 and PV2; b) training a processing unit on the training set TS1 of step a) to identify correlation patterns between the measured one or more process variables and at least one process variable, and developing a calibration function CF1 for the calibrated soft-sensor from the identified correlation patterns; c) predicting at least one operating parameter for the chemical process as an approximation to LV1 and / or PV1; d) calculating the deviation as the difference between the predicted operating parameters and the corresponding laboratory values ​​LV1 and / or process values ​​PV1 of the training set TS1 of step a); e) if the deviation calculated in step d) exceeds a threshold value, proceed with step f), otherwise proceed with step g); f) recalibrating the soft sensor of step b); g) writing the predicted at least one operating parameter to the DCS; h) repeating steps c) through g); The present invention relates to a method, including:

Description

[Technical Field]

[0001] The present invention is in the field of process control, particularly the control of chemical processes, specifically the production of one or more of methanol, hydrogen sulfide, methyl mercaptan, hydrocyanic acid, acrolein, 3-methylthiopropionaldehyde, 5-(2-methyl-mercaptoethyl)-hydantoin, methionine, salts of methionine, and derivatives of methionine. Process control is a combination of the control engineering and chemical engineering disciplines that uses industrial control systems to achieve levels of product consistency, economy, and safety that cannot be achieved through purely manual human control alone. This is widespread in industries such as petroleum refining, pulp and paper manufacturing, chemical processing, and power plants. While ranging greatly in size, variety, and complexity, a small number of operators can manage complex processes with a high degree of consistency. The development of large, automated process control systems has enabled the design of large, high-volume, and complex processes that could not otherwise be operated economically or safely. Applications can range from temperature and level control of a single process vessel to entire chemical processing plants with thousands of control loops. [Background technology]

[0002] Process control in large industrial plants has evolved through many stages. Early on, control was performed from local panels in the processing plant. However, manning these distributed panels required significant manpower and did not provide a complete view of the process. The next logical evolution was to transmit measurements from the entire plant to a permanently manned central control room. Effectively, this was a centralization of all local panels, with the benefits of lower manpower levels and easier process visibility. In many cases, the controller was located behind a control room panel, with all automatic and manual control outputs transmitted back to the plant. However, while providing centralized control, this arrangement offered little flexibility, as each control loop had its own controller hardware and required continuous movement of an operator around the control room to monitor various parts of the process.

[0003] With the advent of electronic processors and graphic displays, it became possible to replace these individual controllers with computer-based algorithms hosted on a network of I / O racks, each with its own control processor, distributed around the plant and able to communicate with graphic displays in one or more control rooms. Thus, the distributed control system was born.

[0004] The introduction of distributed control systems (DCS) allowed for easy interconnection and reconfiguration of plant controls, for example, cascading loops and interlocks, and easy interfacing with other production computer systems. This enabled sophisticated alarm handling, introduced automatic event logging, eliminated the need for physical records such as chart recorders, allowed control racks to be networked and therefore located locally in the plant, reduced cabling, and provided an advanced picture of plant status and production levels.

[0005] Furthermore, the introduction of DCSs enabled the application of more sophisticated control methods that use mathematical optimization to calculate the best operating parameters that meet given constraints. However, this requires a control model to control processes using computers and DCSs, specifically to predict the behavior of the system in question. The control model is a set of equations used to predict the system's behavior and can help determine how it will respond to changes. To determine the basic model for any process, the system's inputs and outputs are defined differently than in other chemical processes. For example, the balance equations are defined in terms of control inputs and outputs rather than material inputs. The state variables (x) are measurable variables that are good indicators of the system's state, such as temperature (energy balance), volume (mass balance), or concentration (component balance). The input variables (u) are specific variables, typically including flow rates.

[0006] Therefore, industrial process plants are typically equipped with numerous sensors. The primary purpose of sensors is to provide data for process monitoring and control. Several decades ago, researchers began to utilize the large amount of data measured and accumulated in process industries by building predictive models based on this data. However, measurement data is still needed to form predictions about the behavior of the system in question. A further problem in this regard is that measuring certain parameters in process plants is often difficult—time-consuming, complex, or even impossible at all. This is where so-called soft sensors come in. Soft sensors, a combination of software and sensor, are virtual sensors. Therefore, they are not actually existing sensors but simulations of the dependence of measurements on representative individual metrics or target values. Therefore, target values ​​are not measured directly, but are calculated or approximated by measurements correlated to the target value and a model of this correlation. At a very general level, two different classes 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 context of the process. However, a significant drawback of such models is that they are primarily developed for the planning and design of process plants and therefore typically focus on describing an ideal steady-state of the process. As a result, such models do not reflect real-world conditions, which are characterized by constantly changing, and especially surprisingly changing, process conditions rather than the ideal steady-state of the process. Therefore, model-driven soft sensors require a significant amount of engineering to accommodate disturbances. This creates significant complexity and may not account for all disturbances, especially when the disturbances are not known in aggregate and the reasons for their occurrence are unknown. This limits the use of model-driven soft sensors for controlling chemical plants.In comparison, data-driven soft sensors are based on data measured within the process plant and therefore describe real-world process conditions more realistically than model-driven soft sensors.

[0007] Despite this, soft sensors, and even data-driven soft sensors, require regular maintenance and adjustment. Maintenance is necessary due to data drift and other changes that degrade the soft sensor's performance. Therefore, soft sensors must be compensated for by adapting or redeveloping the model. However, current soft sensors do not offer any automated mechanisms for their maintenance. Therefore, soft sensors still need to be manually controlled and maintained. And the drawbacks don't end there: there is still no acceptable absolute measure for assessing the quality level of soft sensors. Thus, determining whether a model is performing well still relies on the model operator's subjective perception based on visual interpretation of the deviation between the corrected target value and its prediction. This, in turn, requires extensive experience on the part of the operator. Worse yet, because the operator's judgment is based solely on their subjective perception and not on an objective assessment of higher-level context, the operator cannot be prevented from making erroneous decisions. The situation becomes even more complicated when it comes to chemical processes that involve the multi-step production of chemicals. For example, the production of methionine involves the preparation of several starting and intermediate compounds.

[0008] Therefore, what is needed is a method for enabling automatic control of a chemical process using a soft sensor, where the chemical process is the production of one or more of methanol, hydrogen sulfide, methyl mercaptan, hydrocyanic acid, acrolein, 3-methylthiopropionaldehyde, 5-(2-methyl-mercaptoethyl)-hydantoin, methionine, salts of methionine, and derivatives of methionine, and where the soft sensor is automatically recalibrated.

[0009] It has been found that this problem can be solved by implementing an automatically recalibrating soft sensor in a framework for controlling a chemical process. Specifically, a calibrated soft sensor obtained by training a processing unit on a training set of mutually correlated process values ​​and laboratory values ​​and / or a training set of one group of mutually correlated process values ​​and another group of process values ​​is automatically recalibrated when there is a deviation between the predicted operating parameters for one or more process values ​​that are correlated to the process values ​​and / or laboratory values ​​as approximations to the laboratory or process values ​​and the corresponding laboratory and / or process values. Specifically, in this case, the original training set used to provide the calibrated soft sensor is extended with additional laboratory and / or process values, and the processing unit is trained on the extended training set obtained in this way to obtain a modified calibration function, which is then used to recalibrate the soft sensor. The recalibrated soft sensor obtained in this way is then used to predict process parameters for controlling the chemical process. If no deviation exists between the predicted process parameters and the corresponding laboratory values, the predicted parameters for operating the chemical process are written into the distributed control system (DCS) of the process. Summary of the Invention [Means for solving the problem]

[0010] One subject of the present invention is therefore a method for controlling a chemical process for the production of one or more of methanol, hydrogen sulfide, methyl mercaptan, hydrocyanic acid, acrolein, 3-methylthiopropionaldehyde, 5-(2-methylmercaptoethyl)-hydantoin, methionine, salts of methionine and derivatives of methionine, the method comprising: a) providing a training set TS1, wherein the training set TS1 comprises mutually correlated process values ​​PV1 and PV2, and / or comprises mutually correlated laboratory values ​​LV1 and PV2; b) training a processing unit on the training set TS1 of step a) to identify correlation patterns between the measured one or more process variables and at least one process variable, and developing a calibration function CF1 for the calibrated soft-sensor from the identified correlation patterns; c) predicting at least one operating parameter for the chemical process as an approximation to LV1 and / or PV1, c1) requesting one or more process values ​​corresponding to PV2 from a distributed control system (DCS) of the chemical process; c2) predicting the operating parameters using the calibrated soft-sensor of step b) based on the values ​​of step c1); and d) calculating the deviation as the difference between the operating parameters predicted in step c2) and the corresponding laboratory values ​​LV1 and / or process values ​​PV1 of the training set TS1 of step a); e) if the deviation calculated in step d) exceeds a threshold value, proceed with step f), otherwise proceed with step g); f) recalibrating the soft sensor of step b), f1) extending the training set TS1 of step a) with different laboratory values ​​LV1 and / or different treatment values ​​PV1 or replacing at least a portion of the training set TS1 with different laboratory values ​​LV1 and / or different treatment values ​​PV1 to obtain a training set TS2; f2) training the processing unit of step b) on the expanded training set TS1 of step f1) or on the training set TS2 to correct the calibration function of step b); f3) recalibrating the soft sensor of step b) with the modified calibration function of step f2); f4) returning to step c) using the recalibrated soft sensor of step f3); and g) writing the at least one operating parameter predicted in step c2) into the DCS; h) repeating steps c) through g); and The method includes:

[0011] The method according to the invention is also illustrated in FIG.

[0012] The method according to the present invention does not need to be performed continuously. Rather, it is appropriate to pause after at least one operating parameter predicted in step c2) is written to the DCS in step g). This also allows each system of the chemical process to reach a stable state. Preferably, the method according to the present invention is performed periodically at intervals of 1 to 10 minutes, for example, every 1, 2, 3, 4, 5, 6, 7, 8, 9, or 10 minutes. It is even more preferable that the method be performed periodically at the same time intervals during at least the same phase of the chemical process in question. However, it may make sense to change the time interval when the chemical process in question transitions from one particular phase to another, since such transitions may cause significant changes in the chemical process.

[0013] In the context of the present invention, the term "soft sensor" is used as known to those skilled in the art to describe a predictive model based on measured and accumulated data in the process industry. This term combines the word "software," since the model is usually a computer program, and the word "sensor," since the model and its hardware counterpart deliver similar information. Essentially, 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, in step c) of the method at least one operating parameter for the chemical process is predicted as an approximation to LV1 and / or PV1, so that the predicted value, i.e. the operating parameter predicted in step c), can be compared in step d) with the expected value, i.e. the corresponding laboratory value LV1 and / or process value PV1 of the training set TS1 of 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 process value PV1.

[0015] In particular, step c) includes requesting one or more process values ​​from a distributed control system (DCS) of the chemical process, where the one or more process values ​​correspond to PV2. Thus, the process values ​​requested in step c1) also correlate to the laboratory value LV1 and / or the process value PV1. Therefore, a soft-sensor based on a calibration function CF1 generated for the correlation between the laboratory value LV1 and the process value PV2 and / or the correlation between the process value PV1 and the process value PV2 may be suitable for forming a prediction of at least one operating parameter as an approximation to LV1 and / or PV1.

[0016] Essentially, the method according to the present invention utilizes historical data to extract correlations between process variables. In particular, a processing unit is trained with a training set of historical data, i.e., a training set of values ​​of intercorrelated process variables, to identify correlation patterns between one or more measured process variables and at least one process variable to be predicted or approximated. An equation, i.e., a calibration function, for the soft sensor is then developed from the identified correlation patterns between one or more measured process variables and at least one process variable to be predicted or approximated. To develop the equation, i.e., calibration function, an appropriate training set is required. In particular, a training set TS1 is provided in step a), which includes intercorrelated processed values ​​PV1 and PV2 and / or intercorrelated laboratory values ​​LV1 and PV2. The first step in building the training set is the collection of intercorrelated values, i.e., laboratory value LV1 and processed value PV2, processed value PV1 and processed value PV2, or both of these pairs.

[0017] To properly reflect changes in the progress of the process in question, it is useful to assign appropriate time series to the correlated values. Therefore, each value, i.e., each laboratory value LV1, each process value PV1, and each process value PV2, is assigned a time stencil indicating the time at which the value was recorded and / or the time at which the sample underlying that value was taken. It is also useful to establish a time series for the values. For this purpose, each laboratory value LV1 and / or each process value PV1 is assigned a time pattern extending back in time from the time stencil of LV1 and / or PV1 over a predetermined time span. The present invention is not limited in any way to the predetermined time span. Rather, the predetermined time span is selected taking into account the individual periodic changes in the correlated values. For example, values ​​of the acid excess, which has an influence on the pH value, are collected every 15, 30, 45, or 60 minutes, depending on the individual situation, i.e., the distance between the time at which the acid excess is added, the time at which the pH value should be known, and the resulting distribution of acid in the combined system. Finally, correlation values ​​are linked if the time stencil of the value matches the time pattern of the corresponding correlation value. Specifically, if the time stencil of processed value PV2 matches the time pattern of laboratory value LV1 and / or processed value PV1, processed value PV2 is linked to laboratory value LV1 and / or processed value PV1. For example, processed values ​​with a predetermined time stencil, e.g., 12:15, 12:30, 12:45, etc., match laboratory value LV1, which has a 15-minute time pattern, and are therefore linked.

[0018] In one embodiment, step a) and / or step f1) of the method according to the invention further comprises A1) collecting laboratory values ​​LV1 and treatment values ​​PV2 and / or treatment values ​​PV1 and treatment values ​​PV2; A2) providing each value of step A1) with a time stencil indicating the time at which the value was recorded and / or the time at which the sample underlying said value was taken; A3) assigning a time pattern to each laboratory value LV1 and / or each processed value PV1, going back in time from a time stencil of LV1 and / or PV1 for a predetermined time span; A4) concatenating the laboratory values ​​LV1 and / or the processed values ​​PV1 with one or more processed values ​​PV2 having a time stencil that matches the time pattern of the laboratory values ​​LV1 and / or the time pattern of the processed values ​​PV1; Includes.

[0019] In the context of the present invention, two values ​​are considered to be correlated with each other if one value depends on or is influenced by the other value. An example of a process value that depends on or is influenced by a laboratory value is the dependence of the pH value on the acid excess. Another example of a process value that influences a laboratory value is the dependence of the acid supply on the acid excess. In the context of the present invention, a laboratory value LV1 and / or a process value PV1 to be predicted or approximated for a process value PV2 can be selected. Next, a) a list of tags that depend on or are influenced by the laboratory value, for example, tags for which the pH value depends on the acid excess, or b) a list of tags that influence the laboratory value, for example, tags for which the acid supply influences the acid excess, is created. This procedure allows the collection of laboratory values ​​LV1 that correlate with the process value PV2 and / or process values ​​PV1 that correlate with the process value PV2.

[0020] In another embodiment of the method according to the invention, i) Processed value PV1 and processed value PV2 are considered to be mutually correlated if processed value PV1 depends on or is influenced by processed value PV2, or conversely if processed value PV2 depends on or is influenced by processed value PV1; and / or ii) A laboratory value LV1 and a treatment value PV2 shall be mutually correlated if the laboratory value LV1 depends on or is influenced by the treatment value PV2, or conversely if the treatment value PV2 depends on or is influenced by the laboratory value LV1.

[0021] Specifically, a process value PV2 is collected that allows correlation with the laboratory value LV1 and / or the process value PV1. Therefore, the mutually correlated laboratory value LV1 and the process value PV2 are appropriately selected depending on the chemical process to be controlled. Similarly, the mutually correlated process values ​​PV1 and PV2 are also appropriately selected depending on the chemical process to be controlled. For example, if the laboratory value LV1 or the process value PV1 is related to an excess of acid, the process value PV2 is a pH value. For example, if the laboratory value LV1 or the process value PV1 is related to an ion concentration, the process value PV2 is the ionic conductivity of the respective medium, e.g., the conductivity of a salt of methionine in the respective medium. For example, if the laboratory value LV1 or the process value PV1 relates to the concentration of a particular organic compound to be produced, such as methanol, methyl mercaptan, acrolein, 3-methylthiopropionaldehyde, methionine, or a derivative of methionine, such as the hydroxy analog of methionine, 2-hydroxy-4-(methylthio)butyric acid, then the process value PV2 is the intensity of a characteristic absorption band (by IR, NIR, UV, or Raman spectroscopy) of the organic compound in question. Alternatively, if the laboratory value LV1 or the process value PV1 relates to the concentration of a particular organic compound that is a starting compound for the compound to be produced, such as propene for acrolein, methanol for methyl mercaptan, or hydrogen sulfide for methyl mercaptan, or 5-(2-methylmercaptoethyl)-hydantoin for methionine, then the process value PV2 is the intensity of a characteristic absorption band (by IR, NIR, UV, or Raman spectroscopy) of the organic starting compound in question. For example, if the laboratory value LV1 or the processed value PV1 is the concentration of oxygen, the processed value PV2 will be an oxygen value obtained from an electrochemical measurement, such as a current measurement or a resistance measurement, or from an optical measurement, such as an absorptivity measurement or a fluorescence measurement.

[0022] Preferably, the laboratory value LV1 and / or the process value PV1 are one or more of the group of acid excess, ion concentration, concentration of the particular organic compound to be produced, such as the concentration of methanol, methyl mercaptan, acrolein, 3-methylthiopropionaldehyde, methionine or a derivative of methionine, such as 2-hydroxy-4-(methylthio)butyric acid, and concentration of oxygen, and the process value PV2 is the pH value, ionic conductivity, intensity of a characteristic absorption band (in IR, NIR, UV or Raman spectroscopy) of the organic compound in question and oxygen value obtained from electrochemical measurements, such as amperometric or resistive measurements, or from optical measurements, such as absorbance or fluorescence measurements.

[0023] Each processed value may be a current value, i.e., a value obtained immediately from measurements in a running process, or may be an average value, i.e., a value collected at multiple different times over a period of time and then averaged to form a single value.

[0024] In a further embodiment of the method according to the invention, the processed value is a current value or an average value.

[0025] The use of averaged processed values ​​has the advantage that they are less noisy than the current processed values, and therefore the processed values ​​in the method according to the invention are preferably averaged processed values.

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

[0027] It is also beneficial if the average value satisfies the requirement that the time stencil of the averaged processed value PV2 matches the time pattern of the laboratory values ​​LV1 and / or the processed values ​​PV1. It is therefore preferable to average the aggregated or collected processed values ​​over a predetermined time span in step A4).

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

[0029] Furthermore, the training in step b) of the method according to the present invention can also take into account the residence times of the relevant components, such as the residence times of the starting compounds or compounds to be produced, as mentioned above, relative to the compound to be produced. The residence times can be obtained by measurement or by equations, in particular as the quotient of the reactor volume or device volume to the outlet volumetric flow. This approach makes it possible to take into account the time shift of the chemical reaction.

[0030] In this case, the training procedure in step b) may include the following steps: Step 1: Create a training set TS for a specific period T of historical data with a sampling frequency f_s (e.g., 1 sample every 1-10 minutes). Step 1.1: Consider dwell time in the process of introducing a time delay into the training set Case A: No residence time is considered in the creation of the training set, and therefore no time delay is introduced between the processed values ​​of different process tags. Case B: Consider the residence time and assume that the residence time is constant over time. Case C: A list of process value tags is used to derive residence times by each formula. Residence times are time-dependent due to process dynamics and tank level fluctuations. Step 1.2: Set the period T = T2 - T1, with T1 as the start point and T2 as the end point. Case A: No action is taken Case B: No action is taken · Case C: Read the process value PV2 of the process value tag for T = T2 - T1 - TR, where TR is greater than the assumed maximum residence time or a given residence time in the process. Thereby obtain the set TPVR, where TPVR = PVR(TR), and the sampling points are, according to the selected sampling frequency, the starting point,..., the starting point + TR · Provide a list (PVL) of process tags that correlate with the laboratory value LV, i.e., the target value of the soft sensor, as compiled or obtained heuristically based on a statistical method, e.g., principal component analysis. Collect the laboratory values over time T. The sampling frequency of the laboratory value f_sl is, in most cases, significantly smaller than f_s, e.g., f_sl = 1 / 4 hour. Maintain f_sl << f_s. Read the process history values using the process tags within TR to obtain the set TPVL · Step 1.3: Introduction of time delay · Case A: No action · Case B: The time delay in TPVL is introduced according to the provided constant residence time · Case C: For each timestamp, derive the residence time using TPVR. Introduce the time delay in TPVL according to the derived residence time · Step 1.4: TPVL matches the timestamp of the laboratory value LV. If the time of the laboratory value LVi is t_LVi and the sampling timestamp of TPVL is t_TPVL, then maintain t_TPVL < t_LVi and |t_TPVL - t_LVi| < T_L, assign the laboratory value LV to the sampling of TPVL, and set the measured or derived time gap as T_L. In this way, the number of samples consisting of (TPVL, LV) is T * From f_sl to T_L * T * Extend to f_sl · Step 2: Use TS in step b) of the method

[0031] Each of Cases A to C creates a different training set. Each training set ultimately results in a different soft-sensor. Therefore, the above procedure can create a total of three different training sets or three different soft-sensors.

[0032] The determination and handling of outliers is always a critical point. Outliers can arise due to variations in measurements or as a result of instrumental errors, although the latter can be excluded from a dataset. In general, especially in statistics, outliers are considered data points that differ significantly from other observations but still leave considerable room for subjective interpretation and misinterpretation. On the other hand, including data points on the periphery of a dataset is essential for meaningful and robust calibration and therefore should not be simply skipped simply because they appear strange. This already shows that detecting or identifying outliers is one of the problems, if not the main problem, associated with outliers. This is because there is no strict definition of an outlier. Therefore, ultimately, determining whether an observation is an outlier remains a subjective exercise. Due to the lack of a universally accepted definition of an outlier, various methods exist for detecting outliers. In the context of the present invention, a value is considered an outlier if it has a variance of more than two sigma from its expected value, regardless of whether it is a laboratory value LV1, a processed value PV1, or a processed value PV2. If a measured value has a variance of more than 2 sigma from its expected value, the measurement underlying the value is repeated. In this context, the terms variance, 2 sigma, and expected value are used as they are known in probability theory. Specifically, the term 2 sigma arises from the normal distribution, also known as the Gaussian or Gaussian distribution, where the variance indicated by sigma describes the width of the standard deviation. It holds that approximately 99.45% of all measured values ​​fall within an interval of ±2 sigma deviation from the expected value.

[0033] In another embodiment of the method according to the invention, the measurements underlying the laboratory value LV1, the treatment value PV1 and / or the treatment value PV2 are repeated if the value in question has a variance of more than 2 sigma from its expected value.

[0034] Preferably, the threshold in step e) of the method according to the invention is given by a relative error.

number

number

number

number

number

[0035] In addition to the automatic recalibration of the soft-sensor in step f), it is also possible to trigger such recalibration of the soft-sensor in certain cases. For example, automatic recalibration may be required periodically at predetermined intervals, aperiodically at predetermined stages of the chemical process, or upon change from one stage to another. For example, the predetermined stages of the chemical process may be the start-up phase, continuous or batch operation, and / or shutdown phase of a manufacturing process. An example of a change from one stage to another is a change from continuous operation to batch operation. Other examples of predetermined stages of a chemical process may include individual phases of a purification process, such as the individual phases of separation of a mixture of substances based on the different boiling points of the individual components in a distillation process or the different mobilities of the individual components in a colloidal matrix.

[0036] In an embodiment of the method according to the invention, step f) is also carried out periodically at predetermined intervals or aperiodically at predetermined stages of the chemical process or when changing from one stage to another.

[0037] In addition to or instead of the automatic recalibration of the soft-sensors in step f), manual triggering of the recalibration of the soft-sensors may also be required.

[0038] In an alternative embodiment of the method according to the invention, step f) is triggered whenever it is considered necessary.

[0039] According to the invention, the soft-sensor is recalibrated by a sequence of f1) extending the training set TS1 of step a) by another laboratory value LV1 and / or another processed value PV1.

[0040] During the recalibration of the soft sensor in step f), the training set TS1 of step a) is extended with additional laboratory values ​​LV1 and / or additional process values ​​PV1 to provide a new training set TS2. Thus, step f) further comprises collecting the additional laboratory values ​​LV1, the additional process values ​​PV1, and the additional process values ​​PV2 and linking the respective values ​​that are correlated with each other in order to arrange the values ​​in a time-series framework. All these substeps are performed in the same way as described in steps A1) to A4) above. Thus, the linking of additional values ​​in step A4) also makes it possible to provide a separate training set TS2 that at least partially or completely replaces the training set TS1.

[0041] In another embodiment of the method according to the invention, the training set TS1 is at least partly or completely replaced by a training set TS2 in step f).

[0042] The conditions of chemical processes are subject to constant changes, and in particular the conditions during the start-up phase of a chemical process, e.g., the conditions at the beginning of a chemical reaction carried out in the process, are often significantly different from the conditions during later phases of the chemical process, e.g., generally during a stabilized chemical reaction or process. Therefore, when the training set TS1 is partially replaced by the training set TS2 in step f), it is preferred to replace the part of the training set TS1 representing the oldest conditions of the respective chemical process, i.e., the part of the training set TS1 having the oldest time stencil values, with the most recent values, i.e., the laboratory values ​​LV1 and / or process values ​​PV1 having the current time stencil.

[0043] In a preferred embodiment of the method according to the invention, the laboratory value LV1 and / or processed value PV1 with the oldest time stencil is replaced by the laboratory value LV1 and / or processed value PV1 with the current time stencil.

[0044] Preferably, the laboratory value LV1 and / or the processed value PV1 with the oldest time stencil is at least partially or completely replaced in step f) by the laboratory value LV1 and / or the processed value PV1 with the current time stencil.

[0045] Considering that conditions often differ significantly at certain stages of a chemical process, it is reasonable to at least partially or completely replace the training set TS1 with the training set TS2 at a predetermined stage of the chemical process, such as after the start-up phase or during the shutdown phase. Alternatively, if step f) is performed more frequently than is considered acceptable, i.e., if the number of times step f) is performed exceeds a predetermined threshold within a predetermined period, it is also reasonable to at least partially or completely replace the existing training set TS1 with the training set TS2. In general, the method according to the present invention does not impose any restrictions on the threshold value in step f). Rather, the threshold value is appropriately selected taking into account the framework conditions and requirements of the individual chemical process.

[0046] In a further embodiment of the method according to the invention, the training set TS1 is at least partially or completely replaced by the training set TS2 at a predetermined stage of the chemical process or if the number of executions of step f) within a predetermined period of time exceeds a predetermined threshold.

[0047] According to the present invention, the processing unit is trained with a training set and a calibration function is provided. Typically, the processing unit is an artificial neural network. In the context of the present invention, the term neural network is used synonymously with the term artificial neural network (ANN) to refer to a computing system inspired by, but not identical to, the biological neural networks that make up animal brains. Specifically, an ANN is based on a collection of connected units or nodes called artificial neurons that loosely model the neurons of a biological brain. Each connection can transmit a signal to other neurons, similar to the synapses in a biological brain. After receiving a signal, an artificial neuron can process it and send a signal to connected neurons. The "signals" at the connections are real numbers, and the output of each neuron is calculated by some nonlinear function of the sum of its inputs. Such connections are called edges. Neurons and edges typically have weights that are adjusted as learning progresses. The weights increase or decrease the strength of the signal at the connection. Neurons can have thresholds such that a signal is sent only if the aggregate signal crosses the threshold. Typically, neurons are aggregated into layers. Different layers can perform different transformations on their inputs. A signal travels from the first layer (input layer) to the last layer (output layer), possibly traversing multiple layers. For example, the layer that receives an external input, e.g., the processed value PV2, is the input layer. The layer that forms the final result, i.e., the predicted operating parameters approximated for LV1 and / or PV1, is the output layer. Hidden layers exist between these layers. Multiple connection patterns are possible between the two layers.

[0048] In one embodiment of the method according to the invention, the processing unit is an artificial neural network.

[0049] 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 of the respective chemical process and the requirements to be met.

[0050] In the context of the present invention, the term deep neural network or its abbreviation DNN is used as known to those skilled in the art to refer to an artificial neural network with multiple layers between an input layer and an output layer. Although there are various types of neural networks, 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.

[0051] However, like many artificial neural networks, naively trained DNNs can suffer from many problems. Two common problems are overfitting and computational time. DNNs tend to overfit due to the additional layers of abstraction that can model rare dependencies in the training data. In statistics, overfitting is the formation of an analysis that corresponds too closely or too accurately to a particular data set, which can result in a failure to fit additional data or to reliably predict future observations. An overfitted model is a statistical model that contains more parameters than can be proven correct by the data. The result of overfitting is that some of the residual variation, or noise, is extracted as if it represents the underlying model structure, without being known. In other words, the model obtained in this way memorizes a huge number of examples instead of learning by attending to features. Furthermore, DNNs must consider many training parameters, such as size (number of layers and units per layer), learning rate, and initial weights. Due to the cost of time and computational resources, sweeping through the parameter space for optimal parameters may not be feasible.

[0052] In this case, the use of convolutional neural networks (CNNs) can be helpful. The term convolutional neural network, or its abbreviation CNN, is used as known by those skilled in the art to describe a class of deep neural networks that use a mathematical operation called convolution instead of standard matrix multiplication in at least one of its layers. CNNs are regularized versions of multilayer perceptrons. Multilayer perceptrons are typically fully connected networks, meaning that each neuron in one layer is connected to every neuron in the next layer. The "fully connected" nature of such networks makes them prone to overfitting the data. Typical regularization techniques include adding some form of weight magnitude measure to the loss function. CNNs take a different approach to regularization: they exploit hierarchical patterns in the data and assemble more complex patterns using smaller, simpler patterns. Thus, CNNs are at the lower end of the connectivity and complexity scale. Therefore, CNNs are less prone to overfitting and require fewer computational resources and training data than other artificial neural networks. In the context of the present invention, the need for reduced training data is a major advantage because in chemical processes, where samples are sometimes taken only once per hour, less data is collected than in other fields, yet convolutional neural networks can be used to obtain high-quality predictive models, calibration functions, and soft sensors.

[0053] Therefore, the artificial neural network is preferably a convolutional neural network.

[0054] In the context of the present invention, the term recurrent neural network, or its abbreviation RNN, is used as known to those skilled in the art to describe a class of artificial neural networks in which connections between nodes form a directed graph along a time sequence. This allows recurrent neural networks to exhibit dynamic behavior over time. The term "recurrent neural network" is used interchangeably to refer to two broad classes of networks with similar general structures, one finite impulse and the other infinite impulse. Both classes of networks exhibit dynamic behavior over time. Finite impulse recurrent networks are directed acyclic graphs that can be unrolled and replaced with strictly feedforward neural networks, whereas infinite impulse recurrent networks are directed cyclic graphs that cannot be unrolled. Both finite impulse recurrent networks and infinite impulse recurrent networks can have additional memory states, which can be directly controlled by the neural network.

[0055] According to the present invention, a processing unit is trained on a training set TS1 to generate a calibration soft-sensor based on a calibration function CF1. In the framework of controlling a chemical process using the method of the present invention, the training set TS1 can be formed by a calibration branch (1) that generates a calibration function, including a laboratory management system (LIMS) (2) that provides laboratory values ​​LV1 (3) and a process information management system (PIMS) (5) that provides process values ​​PV1 and / or PV2 (6). For example, a so-called laboratory information management system (LIMS) is polled for laboratory values ​​obtained from experiments performed in a laboratory that represent the process in question. The training set (8) is created by concatenating the laboratory values ​​LV1 with the corresponding process values ​​PV2. Alternatively, the training set can be created by collecting process values ​​PV1 that correlate with at least one process value PV2 and concatenating the process values ​​PV1 with the appropriate process values ​​PV2. By assigning a time stencil to every collected value, it is possible to link the laboratory value LV1 with the appropriate processed value PV2, or the processed value PV1 with the appropriate processed value PV2, which time stencil indicates the time at which each value was recorded and / or the time at which the sample on which it is based was recorded. The training set (8) thus generated is used to train the processing unit (13) to generate a calibration function that allows the processing unit to perform a prediction of the operating parameter, in this case as an approximation to the laboratory value LV1 and / or the processed value PV1.

[0056] According to the present invention, at least one operating parameter is predicted as an approximation to a laboratory value LV1 and / or a process value PV1 of a chemical process having one or more process values ​​corresponding to a process value PV2. In the framework of controlling a chemical process according to the method of the present invention, the one or more process values ​​can be provided by an operating loop (9) requesting process parameter PV2 (12) from a distributed control system (DCS) (10) of the chemical process.

[0057] A processing unit (13) adapted to carry out the method according to the invention is arranged between the calibration branch (1) and the operating loop (9). The processing unit (13) is therefore connected to the calibration branch (1), which supplies it with the laboratory value LV1 and the processed values ​​PV1 and / or PV2 required for generating the calibration function, and to the operating loop (9), which supplies it with the required processed value PV2, so that a prediction of at least one operating parameter is formed as an approximation to LV1 and / or PV1.

[0058] Another subject of the method according to the invention is a system for controlling a chemical process, comprising: i) a calibration branch (1) for generating a calibration function, including a Laboratory Information Management System (LIMS) (2) providing a laboratory value LV1 (3) and a Process Information Management System (PIMS) (5) providing process values ​​PV1 and / or PV2 (6); ii) an operational loop (9) requesting one or more process values ​​(12) from a distributed control system (DCS) (10) of the chemical process; iii) a processing unit (13) adapted to carry out the method according to the invention; wherein the processing unit is connected to a calibration branch (1) and to an operating loop (9).

[0059] Preferably, the system according to the invention is for controlling the production of methanol, hydrogen sulfide, methyl mercaptan, hydrocyanic acid, acrolein, 3-methylthiopropionaldehyde, methionine, salts of methionine, or derivatives of methionine.

[0060] In one embodiment of the device according to the invention, the processing unit is an artificial neural network.

[0061] 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 of the respective chemical process and the requirements to be met. Preferably, the processing unit is a convolutional neural network.

[0062] Optionally, the calibration branch further comprises a laboratory value buffer (4) to which the laboratory values ​​(3) are transferred after they have been recorded. When the laboratory value buffer is full, a corresponding processed value PV2 (5) is collected by a collector (7) from the process information management system (5) for each laboratory value, and a training set is created by concatenating the laboratory value LV1 and / or the processed value PV1 with the processed value PV2 that matches them.

[0063] The predicted operating parameters (14) are then written to a distributed control system (10) of the chemical process by open platform communication (11).

[0064] The present invention will be described below with reference to FIGS. 1 to 4 and examples. [Brief explanation of the drawings]

[0065] [Figure 1] FIG. 1 is a flow diagram of a method according to the present invention. [Figure 2] 1 is a schematic diagram of a method and apparatus according to the present invention, where the individual numbers have the meaning of calibration branch (1), process information management system (PIMS) (2), laboratory values ​​(3), laboratory value buffer (4), process information management system (PIMS) (5), processed values ​​(6), collector (7), training set (8), operating loop (9), distributed control system (DCS) (10), open platform communication (OPC) (11), processed values ​​(12), processing unit (13) and predicted operating parameters (14). [Figure 3] 10 is a graph showing the results of a comparative example. [Figure 4]FIG. 1 is a diagram showing the results of Example 1 according to the present invention. [Figure 5] FIG. 1 is a diagram showing the results of Example 2 according to the present invention. [Figure 6a] FIG. 10 shows the relative error in the prediction of a soft sensor not according to the present invention. [Figure 6b] FIG. 10 illustrates the relative error in prediction of an automatically recalibrating soft sensor according to the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0066] Comparative Example: The sulfuric acid excess in an ammonia scrubber downstream of a reactor for producing hydrogen cyanide was predicted using a soft-sensor without recalibration. Simultaneously, the sulfuric acid excess was measured as an actual value. Figure 3 shows the predicted and actual measurement results, along with the sulfuric acid excess value, one measured in the laboratory (solid black line, laboratory) and one approximated by a prior-art soft-sensor (dotted black line, soft-sensor). The first ellipsis (dashed black line, left) indicates a spike, and the second ellipsis (dashed black line, right) indicates an offset correction. As can be seen from Figure 3, the process value predicted by the soft-sensor approaches the actual measurement. After an initial large deviation, the predicted value follows the trend of the measurement but never matches the measurement. Instead, after a period of synchronization, the predicted value begins to deviate from the actual value, gradually becoming significantly different, and an offset correction must be performed due to the strong discrepancy between the predicted and actual values.

[0067] Example 1 according to the present invention: The sulfuric acid excess in an ammonia scrubber downstream of a reactor for producing hydrogen cyanide was predicted using the method of the present invention. At the same time, the sulfuric acid excess was also measured as an actual value. Figure 4 shows a graph of the predicted and actual measurement results, along with the sulfuric acid excess value, one measured in the laboratory (solid black line, laboratory) and one approximated by the method of the present invention (dotted gray line, prediction). Three ellipses indicate the difference between the approximated and actual values ​​and the immediate correction. As can be seen from Figure 4, the process value predicted by the method of the present invention was in much better agreement with the actual measurement. Furthermore, the method of the present invention was able to identify discrepancies between the predicted and actual values ​​fairly quickly, and the soft sensor could be recalibrated immediately after the deviation was identified, so that very good consistency between the predicted and actual values ​​was restored.

[0068] Example 2 according to the present invention: This example illustrates the recalibration of a soft-sensor in a method according to the present invention. Again, the sulfuric acid excess in an ammonia scrubber downstream of a reactor for producing hydrogen cyanide was predicted using the method according to the present invention and measured in the laboratory. However, in comparison with Example 1, hydrogen cyanide production was shut down and then restarted. After restarting, the soft-sensor calibration function no longer matched the process conditions. As a result, a large offset occurred between the sulfuric acid values ​​measured in the laboratory and the predicted sulfuric acid values. This large offset can be seen in Figure 5 from January 2, 2020, to January 8, 2020. However, once automatic training, i.e., recalibration of the soft-sensor, was initiated (shown by the dotted line), the prediction of sulfuric acid values ​​improved significantly. After January 8, 2020, no offset was observed at all. In the rare cases where a discrepancy occurred between the predicted and actual values, the soft-sensor was automatically recalibrated, and excellent agreement between the predicted and actual values ​​was again achieved. The results are shown in Figure 5 together with the values ​​of the sulfate excess, one measured in the laboratory (cross, laboratory) and one approximated by the method according to the invention (black line, predicted).

[0069] Figures 6a and 6b show the relative error in prediction before recalibration (i.e., before retraining begins) (Figure 6a) and after recalibration (i.e., after retraining) (Figure 6b). Figure 6a shows that a well-trained soft sensor can typically perform predictions with a prediction error of 5% to 35%. However, the relative error before retraining is quite large; for example, after 12 predictions, the relative error is 25% and never reaches zero. Furthermore, the relative error of a soft sensor without automatic recalibration is thought to have some chaotic properties; specifically, its distribution is unbalanced and does not follow a Gaussian distribution.

[0070] By comparison, Figure 6b shows that the auto-recalibrated soft sensor provides an improvement in relative error after retraining, i.e., after recalibration begins, compared to the soft sensor without recalibration. Specifically, the relative error of the predictions for the auto-recalibrated soft sensor is between -10% and +10%, and is therefore much lower in absolute terms. The major improvement is the relative error of zero for many predictions. Furthermore, the error distribution in Figure 6b is well balanced and follows a Gaussian distribution, in contrast to the error distribution in Figure 6a.

Claims

1. 1. A method for controlling a chemical process, the chemical process being the production of one or more of methanol, hydrogen sulfide, methyl mercaptan, hydrocyanic acid, acrolein, 3-methylthiopropionaldehyde, 5-(2-methylmercaptoethyl)-hydantoin, methionine, salts of methionine, and derivatives of methionine, the method comprising: a) providing a training set TS1, said training set TS1 comprising mutually correlated processed values ​​PV1 and PV2 and / or mutually correlated laboratory values ​​LV1 and PV2; b) training a processing unit on the training set TS1 of step a) to identify correlation patterns between the measured process variable(s) and at least one process variable, and developing a calibration function CF1 for the calibrated soft-sensor from the identified correlation patterns; c) predicting at least one operating parameter for the chemical process as an approximation to LV1 and / or PV1, c1) requesting one or more process values ​​corresponding to PV2 from a distributed control system (DCS) of the chemical process; c2) predicting the operating parameters with the values ​​of step c1) by the calibrated soft-sensors of step b); and d) calculating the deviation as the difference between the operating parameters predicted in step c2) and the corresponding laboratory values ​​LV1 and / or processed values ​​PV1 of the training set TS1 of step a); e) if the deviation calculated in step d) exceeds a threshold value, proceed with step f), otherwise proceed with step g); f) recalibrating the soft sensor of step b), f1) extending the training set TS1 of step a) with other laboratory values ​​LV1 and / or other process values ​​PV1 or replacing at least a part of the training set TS1 with other laboratory values ​​LV1 and / or other process values ​​PV1 to obtain a training set TS2; f2) training the processing unit of step b) on the expanded training set TS1 or training set TS2 of step f1) to modify the calibration function of step b); f3) recalibrating the soft-sensor of step b) with the modified calibration function of step f2); f4) returning to step c) using the recalibrated soft sensor of step f3); and g) writing the at least one operating parameter predicted in step c2) into the DCS; h) repeating steps c) through g); and A method comprising:

2. Step a) and / or step f1) A1) collecting laboratory values ​​LV1 and processed values ​​PV2 and / or processed values ​​PV1 and PV2; A2) providing each value of step A1) with a time stencil indicating the time at which the value was recorded and / or the time at which the sample underlying said value was taken; A3) applying a time pattern to each laboratory value LV1 and / or each processed value PV1, going back in time from a time stencil of LV1 and / or PV1 for a predetermined time span; A4) concatenating the laboratory values ​​LV1 and / or the processed values ​​PV1 with one or more processed values ​​PV2 having a time stencil that matches the time pattern of said laboratory values ​​LV1 and / or the time pattern of said processed values ​​PV1; further comprising: The method of claim 1.

3. i) processed values ​​PV1 and PV2 are considered to be correlated with each other if processed value PV1 depends on or is influenced by processed value PV2, or conversely, processed value PV2 depends on or is influenced by processed value PV1; and / or ii) A laboratory value LV1 and a processed value PV2 are mutually correlated if the laboratory value LV1 depends on or is influenced by the processed value PV2, or conversely if the processed value PV2 depends on or is influenced by the laboratory value LV1; 3. The method according to claim 1 or 2.

4. The processed value is a current value or an average value.

4. The method according to any one of claims 1 to 3.

5. the average value is obtained by averaging the aggregated processed values ​​over a predetermined time span; The method of claim 4.

6. the average value is obtained by averaging the aggregated processed values ​​over a predetermined time span of step A4). The method of claim 4.

7. The measurements underlying the laboratory value LV1, the processed value PV1 and / or the processed value PV2 are repeated if the value in question has a variance of more than 2 sigma from its expected value; 7. The method according to any one of claims 1 to 6.

8. Step f) may also be performed periodically at predetermined intervals or aperiodically at predetermined stages of the chemical process or upon change from one stage to another, 8. The method according to any one of claims 1 to 7.

9. Step f) is triggered whenever it is considered necessary; 9. The method according to any one of claims 1 to 8.

10. The training set TS1 is at least partially or completely replaced by the training set TS2 in step f), 10. The method according to any one of claims 1 to 9.

11. The laboratory value LV1 and / or the processed value PV1 with the oldest time stencil are replaced by the laboratory value LV1 and / or the processed value PV1 with the current time stencil; The method of claim 10.

12. at a predetermined stage of the chemical process or when the number of executions of step f) within a predetermined period of time exceeds a predetermined threshold, the training set TS1 is at least partially or completely replaced by the training set TS2; 12. The method according to any one of claims 1 to 11.

13. the processing unit is an artificial neural network; 13. The method according to any one of claims 1 to 12.

14. The artificial neural network is a convolutional neural network.

14. The method of claim 13.

15. 1. A system for controlling a chemical process, the system comprising: i) a calibration branch (1) for generating a calibration function, including a Laboratory Information Management System (LIMS) (2) providing a Laboratory Value LV1 (3) and a Process Information Management System (PIMS) (5) providing Process Values ​​PV1 and / or PV2 (6); ii) an operational loop (9) for requesting one or more process values ​​(12) from a distributed control system (DCS) (10) of said chemical process; iii) a processing unit (13) adapted to carry out the method according to any one of claims 1 to 14 and connected to the calibration branch (1) and to the operating loop (9); A system with.

Images