Predicting values of a process variable

Abstract

The invention proposes a method for the computer-assisted prediction of values of a first process variable, comprising: automated training of a computer-implemented model, which comprises relationships between the captured process measurement values of the second process variable and the values to be predicted of the first process variable, using the previously captured process measurement values, the following steps being carried out in an automated manner in the course of the training: i) scaling the process measurement values, in particular standardizing the process measurement values, ii) carrying out a data imputation in order to supplement missing process measurement values, iii) detecting and removing outliers in the process measurement values, iv) performing a feature selection in the process measurement values, v) increasing a quantity of the process measurement values, vi) using a specific window size for the process measurement values when training the computer-implemented model, vii) selecting a specific type for the computer-implemented model and hyperparameters for training the computer-implemented model, predicting the values of the first process variable by applying the previously trained computer-implemented model.

Description

[0001] 202419537 Foreign version 03.11.2025

[0002] 1

[0003] Description

[0004] Prediction of values ​​of a process variable

[0005] The invention relates to a method for the computer-aided prediction of values ​​of a process variable, having the features of claim 1. Furthermore, the invention relates to a system for the computer-aided prediction of values ​​of a measured variable according to claim 4.

[0006] To control a process plant, a distributed control system (DCS) or a SCADA system (Supervisory Control and Data Acquisition System) collects measured values ​​and controls actuators such as valves, pumps, etc. Multiple controllers, e.g., PID or MPC, control the process. Field devices for in-situ data acquisition can be expensive (CAPEX and / or OPEX), or in-situ technology may not be available. If no in-situ technology is available, laboratory samples are taken and evaluated. The results are typically available with a significant time delay, making them unsuitable for use by the controller. To overcome this challenge, data-driven, stringent, or hybrid soft sensors are used to predict process values ​​and make them available, for example, for plant control. Other applications include quality prediction, reduced energy consumption, or a customer-specific optimization criterion.

[0007] A data-driven soft sensor should be trained with minimal manual intervention, without requiring expert knowledge; that is, neither a data scientist nor a subject matter expert should be involved. Furthermore, it should be able to predict both a value and the quality of the prediction (uncertainty estimation).

[0008] Data-driven soft sensors utilize various machine learning techniques to model and analyze datasets. This task is typically performed by an experienced data scientist in collaboration with a subject matter expert.

[0009] This approach has the disadvantage of being expensive, requiring a data scientist and usually a subject matter expert, and the outcome is highly dependent on the expertise of both specialists and their ability to collaborate and communicate. 202419537 Foreign version 03.11.2025

[0010] 2

[0011] Furthermore, a new model generation approach is currently developed for each new project, as a successful ML model cannot be transferred to another application. Proven AutoML techniques can be used for the development of soft sensors, but require adaptations.

[0012] More sophisticated techniques utilize automated machine learning (AutoML). However, well-known AutoML pipelines such as Auto-Sklearn, TPOT, or H2O AutoML have limitations regarding their feature space and applicability to other domains. This means they may not cover all the techniques required for developing a data-driven soft sensor in the process industry. Therefore, collaboration with data scientists remains essential to achieve a high-quality result.

[0013] The predictive accuracy of a data-driven or hybrid soft sensor can decrease over time due to slow changes in the equipment or changes in the input materials.

[0014] For strict soft sensors, a Kalman filter can be used. This approach has the disadvantage that the underlying physics of the process must be understood and modeled. If the model does not already include the changing process behavior, which is the standard case, the model must be extended or new process parameters estimated. For example, the inclusion of laboratory results in the online estimation of a Kalman filter to improve its estimation results is disclosed in EP 2544056 A1. Transferring this method to data-driven soft sensors is not possible.

[0015] In “LI LEI-JUN ET AL: "Prediction confidence-based dynamic selection and weighted integration", CONCURRENCY AND COMPUTATION: PRACTICE AND EXPERIENCE, Vol. 33, No. 8, 28 November 2018 (2018-11-28), XP093289771 , GB ISSN: 1532-0626, DOI: 10.1002 / cpe.5055” a system for the computer-aided prediction of values ​​of a process variable of an industrial process is disclosed.

[0016] For data-driven soft sensors, the machine learning model can be retrained. This approach has the disadvantages of high computational costs, slow adaptation of the model to the new data points, and no measure to address the uncertainty of the laboratory sample. 202419537 Foreign version 03.11.2025

[0017] 3

[0018] The invention is based on the objective of providing a method and an associated device that overcome the aforementioned disadvantages and enable an efficient and accurate prediction of values ​​of a process variable.

[0019] The problem stated above is solved by a method for computer-aided prediction of the values ​​of a process variable with the features of claim 1. Furthermore, the problem stated above is solved by a system for computer-aided prediction of the values ​​of a process variable according to claim 4. Advantageous embodiments are the subject of the dependent claims.

[0020] A method according to the invention for the computer-aided prediction of values ​​of a first process variable comprises the following process steps: a) carrying out an industrial process in an industrial plant, in particular a manufacturing plant or process plant, b) acquiring process measurements of at least a second process variable of the industrial process by at least one physical sensor, c) automating the training of a computer-implemented model, which includes relationships between the acquired process measurements and the values ​​of the first process variable to be predicted, with the previously acquired process measurements, wherein the following steps are carried out automatically during the training: i) scaling of the process measurements, in particular standardization of the process measurements, ii) performing data imputation to supplement missing process measurements, iii) detecting and removing outliers in the process measurements.iv) Performing feature selection in the process measurements, v) Performing an augmentation of a set of process measurements, vi) Using a specific window size for the process measurements when training the computer-implemented model, vii) Selecting a specific type for the computer-implemented model and hyperparameters for training the computer-implemented model, d) wherein the automated training is performed using a Combined Algorithm Selection and Hyperparameter (CASH) optimization, under tuning by a Bayesian optimization, to generate a plurality of possible models, estimating an uncertainty in the prediction of the values ​​of the first process variable. 202419537 Foreign version 03.11.2025

[0021] 4. The model with the lowest uncertainty is used to predict the values ​​of the first process variable. e) Recalibrating the prediction of the values ​​of the first process variable using metrologically determined values ​​of the first process variable, preferably using laboratory samples, wherein performing the recalibration comprises the following steps: i) assigning weights to several models of the plurality of models; ii) updating the weights based on comparisons between the predictions of the values ​​of the first process variable of each model and the metrologically determined values ​​of the first process variable or the values ​​of the first process variable using the laboratory samples; iii) performing the prediction based on the updated weights of the models.

[0022] An "industrial process" generally refers to the industrially operated transformation of raw materials into finished products. These processes can encompass various methods to alter the properties of the raw materials, such as size, shape, density, or color.

[0023] The industrial plant in question can be a plant from the process industry, such as a chemical, pharmaceutical, petrochemical, or food and beverage plant. This also includes any plant from the manufacturing industry, such as factories where cars or goods of all kinds are produced. Technical plants suitable for carrying out the process according to the invention can also originate from the energy generation sector. Wind turbines, solar power plants, or power stations for energy generation are likewise included in the term "technical plant."

[0024] These systems each have a control system or at least a computer-aided module for controlling and regulating the ongoing process or production. Part of the control system or control module, or of the technical system, is at least a database or archive in which historical data is stored.

[0025] It is initially assumed that the industrial process is in operation, with at least one physical sensor transmitting process measurements from a (second, from the 202419537 foreign version 03.11.2025)

[0026] The first five different process variables of the industrial process are measured and made available for further processing. A physical sensor, in this context, refers to a real sensor actually present in the industrial plant (as opposed to "virtual soft sensors").

[0027] These process measurements of the (second) process variable are automatically used to train a computer-implemented model. Machine learning is therefore employed to generate a suitable model for predicting the values ​​of the first process variable. This model incorporates relationships between the acquired process measurements and the predicted values ​​of the measured variable. During the training process, the following steps are performed automatically, i.e., without the involvement of a data technician, operator, or similarly trained person: i) Scaling of the process measurements, in particular standardization of the process measurements: Scaling the process measurements during model training refers to the process of adjusting the characteristics of the process measurements to a consistent range.This represents an important processing step aimed at normalizing the process measurements and converting them into a standardized format. ii) Performing data imputation to fill in missing process measurements: Data imputation involves replacing missing data points in the process measurements to improve analysis accuracy and avoid biases. This is important when training machine learning models, as incomplete data can significantly impair model performance. iii) Detecting and removing outliers in the process measurements: Outliers are data points that deviate significantly from other observations and can negatively impact model performance.iv) Performing feature selection in process measurements: Performing feature selection in machine learning refers to the process of identifying and selecting the most relevant features (or variables) from the process measurements to improve the efficiency and accuracy of a model. This 202419537 foreign version 03.11.2025.

[0028] 6

[0029] This step is important to reduce overfitting, shorten training time, and increase the interpretability of the model. v) Performing an increase in the set of process measurements: This step is also known as data augmentation and refers to techniques for artificially increasing the amount of training data by generating new data points from the existing process measurements. This is particularly useful for improving the robustness and generalizability of the model, especially if the original set of process measurements is too small. vi) Using a specific window size for the process measurements when training the computer-implemented model: Using a specific window size for the process measurements in machine learning refers to the "sliding window" or "rolling window" technique.This method is frequently used in the processing of time series data to identify patterns and trends over defined time intervals. vii) Selecting a specific type for the computer-implemented model and hyperparameters for training the computer-implemented model: The most suitable model type is automatically selected, with options such as linear model, tree model, support vector machine, Bayesian model, or K-nearest neighbors. Hyperparameters are parameters that are set before training the model and can significantly influence model performance. Examples of such hyperparameters include a learning rate, a number of sends, a number of neurons, a regulation, a number of trees, or core functions.

[0030] After training the model, it is used to predict values ​​of the first process variable. This application of the model can preferably be carried out in parallel with the operation of the industrial plant in order to use the determined values ​​of the first process variable for controlling the industrial process. The model can also be used to identify potential critical situations during the execution of the industrial process in advance and to initiate countermeasures in a timely manner. Furthermore, the model can be used to predict the maintenance requirements of individual components of the industrial plant. 202419537 Foreign version 03.11.2025

[0031] 7

[0032] Automated training is performed using Combined Algorithm Selection and Hyperparameter (CASH) optimization, refined through Bayesian optimization to generate a plurality of possible models. The uncertainty in predicting process variable values ​​is estimated, and the model with the lowest uncertainty is selected. Within the CASH optimization process, the best algorithm and its optimal hyperparameters are simultaneously chosen for the model to be trained. This presents an optimization problem aimed at maximizing the model's performance by optimally adjusting both the algorithm and its hyperparameters. Bayesian optimization is employed to automate the process.

[0033] As part of this CASH optimization, a plurality of possible models are generated. For each of these models, the uncertainty of the prediction of the values ​​of the first process variable is estimated, and the model with the lowest uncertainty is used for the valid prediction of the values ​​of the first process variable.

[0034] To estimate uncertainty, an ensemble technique can be used to generate a confidence interval. This involves using a variety of generated models to create (even) more robust and accurate confidence intervals, thus obtaining the best-fitting model for predicting the first process variable with greater certainty. Possible ensemble techniques include bootstrap aggregation, random forests, stacking, or Bayesian model averaging.

[0035] The prediction of the first process variable is recalibrated using measured values ​​of that variable. This recalibration can be performed once or repeatedly, preferably periodically. The values ​​of the first process variable are measured directly using a physical sensor, often in a complex and expensive process. These actual measured values ​​are then used to recalibrate the prediction. If direct measurement in the industrial process is not possible or is too costly, values ​​from laboratory samples can also be used for recalibration.

[0036] Performing the recalibration includes the following steps: a) Assigning weights to several models of the majority of models; 202419537 Foreign version 03.11.2025

[0037] 8 b) Updating the weights based on comparisons between the predictions of the values ​​of the first process variable of each model and the measured values ​​of the first process variable or the values ​​of the first process variable using the laboratory samples; c) Performing the prediction based on the updated weights of the models.

[0038] Compared to previously known methods, the described update of the model weights can be carried out particularly efficiently and in a resource-saving manner, which allows the recalibration to be performed with minimal effort, thus significantly increasing its applicability.

[0039] Updating the weights can involve increasing weights for models with predictions that are closer in magnitude to the measured values ​​or laboratory samples, and decreasing weights for models with predictions that are further away from the measured values ​​or laboratory samples.

[0040] The previously formulated task is also solved by a system for the computer-aided prediction of values ​​of a first process variable, comprising a processor and a memory containing instructions which, when executed by the processor, cause the system to carry out a procedure as previously described. The system can be part of a control system for the industrial plant and, in particular, include an operator station server on which the system is implemented. In this context, an "operator station server" is understood to be a server that centrally collects data from an operator control and monitoring system, as well as, typically, alarm and measurement data archives from a control system of an industrial plant, and makes this data available to users.The Operator Station Server typically establishes a communication link to the automation systems of the industrial plant and forwards data from the technical system to so-called clients, which are used to operate and monitor the individual functional elements of the industrial plant. The Operator Station Server can have client functions to access the data (archives, messages, tags, variables) of other Operator Station Servers. This allows images of the industrial plant's operation on the Operator Station Server to be combined with variables from other Operator Station Servers (server-to-server communication). (Operator Station 202419537 Foreign Version 03.11.2025.)

[0041] 9

[0042] The server could, without being limited to, be a SIMATIC PCS 7 Industrial Workstation Server from SIEMENS.

[0043] The properties, features and advantages of this invention described above, as well as the manner in which they are achieved, will become clearer and more easily understood in connection with the following description of the exemplary embodiment, which is explained in more detail in conjunction with the drawing.

[0044] The figure schematically illustrates a method according to the invention. In a first step (I), an industrial process such as combustion in a gas turbine is carried out. The first process variable, a temperature within the gas turbine, is to be predicted by a model. For this purpose, in a second step (II), an ambient temperature of the industrial process is measured as the second process variable using a physical sensor, and the determined process measurements are supplied to an operator station server as part of a control system of the industrial plant. The operator station server comprises a processor and a memory containing specific instructions for the automated execution of the following process steps (III-X):

[0045] In the first subsequent step III, the determined process measurements are scaled. In the subsequent step IV, data imputation is performed to supplement missing process measurements. In the subsequent step V, outliers in the process measurements are identified and removed. In the subsequent step VI, feature selection is performed within the process measurements. In the subsequent step VII, a set of process measurements is enlarged. In the subsequent step VIII, a specific window size for the process measurements is used when training the computer-implemented model. In the subsequent step IX, a specific type of computer-implemented model and hyperparameters for training the computer-implemented model are selected.In a subsequent step X, the values ​​of the first process variable are finally predicted by applying the previously trained computer-implemented model.

[0046] In a final step XI, the predicted values ​​of the first process variable, i.e., the temperature in the turbine, are used to control the combustion in the turbine via the control system. 202419537 Foreign version 03.11.2025

[0047] 10

[0048] According to one embodiment of a method according to the invention, automated training is performed using Combined Algorithm Selection and Hyperparameter (CASH) optimization, with fine-tuning by Bayesian optimization. This involves recalibrating the prediction of the values ​​of the first process variable using metrologically determined values ​​of the first process variable. Performing the recalibration comprises the following steps: First, weights are assigned to several models of the plurality of models. t ' = ^=i Wi * y t ,i

[0049] With y t 'This refers to the prediction of the values ​​of the first process variable. This is derived from the respective product of weight and the model y t i , summed over the plurality n of models.

[0050] Initially, all particles are weighted equally (e.g., 14! = w2 = w). nWith each new (real-world) process measurement for the first process variable (or with each new laboratory sample), the particle filters (weights) are updated by comparing the prediction of each model in the ensemble with the current laboratory sample or the current real-world process measurement. If a model has a small error with respect to the laboratory sample / real-world process measurement (i.e., is "close" to the laboratory sample / real-world process measurement), the model's weight in the ensemble is increased. If a model has a large error with respect to the laboratory sample / real-world process measurement (i.e., is "far" from" the laboratory sample / real-world process measurement), the model's weight in the ensemble is decreased.

[0051] Wi = Wt * pe t ,i)

[0052] Here, p(e) denotes the probability of each model's error. If all particles except one have a weight of 0, meaning they are inactive, the following steps can be used:

[0053] Is there a model in the ensemble that has a lower error with respect to the laboratory sample / the real process measurement? argmax(w #= argmin(e t )i ? 202419537 Foreign version 03.11.2025

[0054] 11

[0055] If so, resampling is performed based on the current error probability: (for each

[0056] If not, the prediction will be made based on the current weights.

[0057] Although the invention has been illustrated and described in detail by the preferred embodiment and the figures, the invention is not limited by the disclosed examples and other variations can be derived from them by the person skilled in the art without leaving the scope of protection of the invention.

Claims

202419537 Foreign version 03.11.2025 12 Patent claims 1. A method for computer-aided prediction of values ​​of a first process variable, comprising: a) carrying out an industrial process in an industrial plant, in particular a manufacturing plant or process plant, b) acquiring process measurements of at least a second process variable of the industrial process by means of at least one physical sensor, c) automating the training of a computer-implemented model, which includes relationships between the acquired process measurements of the second process variable and the values ​​of the first process variable to be predicted, with the previously acquired process measurements, wherein the following steps are carried out automatically during the training: i) scaling of the process measurements, in particular standardization of the process measurements, ii) performing data imputation to supplement missing process measurements, iii) detecting and removing outliers in the process measurements.iv) Performing feature selection in the process measurements, v) Performing an increase in the size of a set of process measurements, vi) Using a specific window size for the process measurements when training the computer-implemented model, vii) Selecting a specific type for the computer-implemented model and of hyperparameters for training the computer-implemented model, d) wherein the automated training is performed using a Combined Algorithm Selection and Hyperparameter (CASH) optimization, under tuning by Bayesian optimization, to generate a plurality of possible models,wherein an uncertainty in the prediction of the values ​​of the first process variable is estimated and the model with the lowest uncertainty is used for a prediction of the values ​​of the first process variable. e) Recalibrating the prediction of the values ​​of the first process variable using metrologically determined values ​​of the first process variable, preferably using laboratory samples, wherein performing the recalibration comprises the following steps: 202419537 Foreign version 03.11.2025 13 i) Assigning weights to several models of the plurality of models; ii) Updating the weights based on comparisons between the predictions of the values ​​of the first process variable of each model and the metrologically determined values ​​of the first process variable or the values ​​of the first process variable using the laboratory samples; iii) Performing the prediction based on the updated weights of the models.

2. The method of claim 1, wherein an ensemble technique is used to generate a confidence interval when estimating the uncertainty.

3. Method according to claim 1 or 2, wherein updating the weights comprises increasing weights for models with predictions that are closer in magnitude to the measured values ​​of the first process variable or closer to the laboratory samples, and decreasing weights for models with predictions that are further away from the measured values ​​of the first process variable or from the laboratory samples.

4. System for computer-aided prediction of values ​​of a first process variable, comprising a processor and a memory containing instructions which, when executed by the processor, cause the system to perform a method according to any one of claims 1 to 3.

Images