TW202409756A - Advanced analytics for bio-/chemical and semiconductor production sites - Google Patents

Abstract

A method to improve bio-/chemical production process performed on a production site (7) by at least one computer (4, 5) the following steps of Gathering and preparing data (2) about the production process in form of process variables by the at least one computer (4, 5); Optimizing and scaling, depending on the process variables, a suitable Machine Learning Model (3) by the at least one computer (4, 5); Creating statistical univariate and/or multivariate profiles of a unit operation of the production site (7) via the optimized Machine Learning Model (3) by the at least one computer (4, 5); Monitoring at least one online process variable of the production process via a digital dashboard (10) provided by the at least one computer (4, 5) by applying the created statistical univariate and/or multivariate profiles to it to detect anomalies, disruptions and/or malfunctions of the production process; and Adapting the production process by either correcting the detected anomalies, disruptions and/or malfunctions and/or indicating the user (9) to interrupt the production process if required quality standards are breached, comprised/violated

Description

For advanced analysis of biological/chemical and semiconductor production sites

The present invention relates to the technical field of AI-enhanced biological/chemical and semiconductor production processes.

The invention described herein discloses a method and system for improving biological/chemical and semiconductor (possibly any) manufacturing processes by applying advanced analytical data-driven AI models.

In line with Industry 4.0 and digitalization, innovative factory technologies exist in the chemical, pharmaceutical and semiconductor industries to increase productivity, efficiency and reduce costs and time to market. Modular chemical plants are deployed to solve these given industrial problems due to the standardization of production equipment. Such plants are being used in industry, but they do not bring all the benefits of Industry 4.0. Unfortunately, data science and automation technology have developed in different directions and now there is a gap between the two. This applies not only to modular chemical plants, but to any plant. Modular production plants can be quickly generated by assembling these pre-engineered modular equipment for a specific process rather than hard programming an entire plant for selected multiple processes. A subject matter expert in a laboratory can build his own modular chemical plant without having to individually program production equipment and drive his processes. These processes are primarily used to experiment with a reaction along with its raw material input data and its final quality data before scaling up to pilot or production scale. The same reaction will be driven frequently. Of course, during these experiments, each attempt will be different from the other due to various sensor measurements and raw material characteristics. Some of these attempts can have drastic variations. The subject matter expert needs to optimize his final quality output data and reduce the quality fluctuations of the sensor data and raw material inputs he has on hand. Even using innovative production methods does not directly provide an option to optimize the process. To optimize the process, sensor data, raw material input data, and final quality data must be manually collected by the subject matter expert and sent to a data science team to interpret the model and possible improvement suggestions. Since correlation does not imply causation, the data science team must have in-depth chemistry and process knowledge to draw scientifically sound conclusions. For a process with only machine learning knowledge, the subject matter expert can rationalize the results better than an unfamiliar colleague. Therefore, allowing the subject matter expert to take ownership of the analysis saves a step and ensures that process knowledge is incorporated into the analysis technique.

There is also no possibility to monitor the process in real time and compare it to the gold trajectory established from the selected best batch or to monitor the univariate profile of gold from the best batch in a time series manner using basic univariate or multivariate statistical procedures.

Taking into account the arguments given above, there is also no technical capability to instantly monitor the process or alert subject matter experts in the event of a failure or interruption during a process. To account for any faults, interruptions, failures, etc. during a process, a model may be utilized, such as a data-driven, entity, or hybrid model.

The scope of this application is for bases with a sufficient level of digitization, such as modular chemical plants. In the absence of digitized modules or plants, this application can still be used by uploading data or by mapping another storage system or using any existing communication interface to the module.

As mentioned above, even identical procedures differ from each other due to natural processes. Sometimes, a step within a complex program with multiple steps can change due to a new raw material batch, a new raw material source, or an updated module/factory. In this case, it must be verified whether the program has changed significantly since it was changed from a previously verified program. In addition, identifying any technical features of similar programs can help avoid deviations in final quality.

Problem: 1-) Subject matter experts cannot independently build data-driven models or use statistical methods to optimize processes at laboratory, pilot, and production scales. The goals of these models and methods are diverse, such as understanding the root cause of a failed product, comparing multiple products to select products with less variation, and identifying processes that are most similar to selected historical processes. 2-) Subject matter experts must manually collect their process data, combine the process data with their raw material and quality data, and send it to their data science team for analysis. This significantly increases the time required for analysis and delays the quality control process. 3-) Even the latest production technologies do not give the subject matter expert the possibility to create univariate and multivariate profiles or golden tracks and use them to monitor these profiles in a time series manner during an ongoing process. 4-) There is no possibility to warn the subject matter expert before or during interruptions or failures during a process. Furthermore, there is no mechanism to autonomously drive a process by using a model that simultaneously monitors and controls each process parameter. 5-) In case of changes in process steps (such as new raw materials or new plants/modules), the final product quality can change. There is no mechanism to verify such changes, whether they are within our target boundaries, and finally to create a certificate of the results. 6-) There is no mechanism to support the selection of the most similar process to compare selected approved historical process data. Another application will identify drift in processes and help users perform predictive maintenance.

Therefore, the mission of this patent application is to disclose a method for improving the efficiency and productivity of digitalized bio/chemical and semiconductor production processes.

This task has been solved by a method for improving biological/chemical and semiconductor production processes performed at a production base, comprising the following steps: collecting/uploading and preparing data about the production process in the form of process variables by the at least one computer; optimizing and calibrating a suitable machine learning model by the at least one computer depending on the process variables; establishing by the at least one computer, via the optimized machine learning model, a statistical unit variable of a process operation of the production base and/or or multivariate profiles; monitoring at least one online process variable of the production process via a digital instrument panel provided by the at least one computer by applying the established statistical univariate and/or multivariate profiles to at least one online process variable to detect anomalies, interruptions and/or failures of the production process; and adjusting the production process by correcting the detected anomalies, interruptions and/or failures and/or instructing the user to interrupt the production process if the required quality standards are violated (or included). Using only one online process variable thereby results in a univariate profile, while using several online process variables results in a respective multivariate profile, whereby the multiple online process parameters are integrated into a multivariate score, which is then monitored and used for the multivariate profiles. Process variables may be, for example, specific sensor inputs, such as a weight or temperature sensor from one of the machines monitoring the production site. The steps of building a data-driven model and interpreting plots as described can be automated on a digital and interactive dashboard to collect and train a model without having to know anything about machine learning or perform any physical labor. The following data-driven steps are automated: data collection, data mapping, data cleaning, data transformation, data calibration, generating a model using machine learning methods, performing hyperparameter tuning on the generated model, and running functions to interpret the model. Given these automated steps, the subject matter expert need only select their projects, batches, and/or variables to analyze to build the model data. After the user inputs, an algorithm and calibrator can be selected to analyze the data. This algorithm and calibrator can be predefined to perform the entire analysis process only. After the "Generate Model" button, a model will be fit, optimized and interpreted. The results of this interpretation will be displayed on the dashboard. The dashboard further allows its users to create statistical univariate and multivariate profiles of a process step, such as a Golden Trail. Once the subject matter expert selects the corresponding function and the project-batch-variable information, these univariate and multivariate profiles are automatically calculated. These created univariate and/or multivariate profiles can then be used to monitor the online process data during a reaction period and compare it to the Golden Trail or keep the current process step between statistical thresholds. This single and multi-variable online monitoring will improve process quality by detecting anomalies, interruptions, failures and help identify drift in the process. This can be used as an early warning system for subject matter experts. Given a quantitative input indicating the performance of a product and historical data about the product, the dashboard allows the user to predict the performance of a batch currently being produced or recently produced. Ultimately, the user can be provided with detected anomalies, interruptions and/or failures to improve the production process by correcting such issues. Furthermore, if the detected issue is so serious that the production process needs to be stopped immediately, the software informs the user via the dashboard so that the user can take appropriate action or at least allow the process to be automatically stopped. These issues can then be resolved permanently or the production process can be stopped completely if necessary. In addition, these data sources can be used to monitor each step of a process on a control chart. Due to the mapped/uploaded data, if steps of a process are out of control, they can be controlled by checking for outliers, deviations, trends and drifts.

Advantageous and therefore preferred further developments of the invention emerge from the associated sub-technical solutions and from the description and the associated drawings.

Preferred further developments of the method of the invention include (for example but not limited to): 1. The machine learning model is established by analyzing the process variables via a first algorithm, wherein the analysis results are displayed to a user via the digital dashboard, then the user selects a suitable model type depending on the analysis results, and then the selected model is established and optimized by the computer by applying a suitable machine learning method (such as principal least squares regression, random forest regression, XGBoost or principal component analysis (PCA)). 2. The user selects at least one batch for the production process as the source of the online process variables and also selects specific process variables to be monitored via the digital dashboard. 3. For statistical univariate profiles, at least two different batches and at least one process variable are selected to calculate statistical features, such as mean, ±2σ, and ±3σ (which can vary between +0.5σ and +7σ). 4. An existing suitable machine learning model is used by the first algorithm by analyzing the process variables, wherein depending on the analysis results, a suitable model is imported and optimized, which can completely drive the process on a given optimization golden track without any further user input. 5. The first algorithm uses several interpretation functions (such as permutation feature importance, Shapley additivity explanation (SHAP) or Gini importance) to find and analyze basic process variables. 6. The machine learning model additionally includes a physical model leading to a mixed model. 7. The machine learning model uses a specific second algorithm, such as principal component analysis (PCA), partial least squares (PLS), random forest, extreme gradient boosting (XGBoost), support vector machine, Gaussian program, multi-layer perceptron, which can be linear, nonlinear or a neural network machine learning method, and uses each second algorithm to complete a regression analysis and provide the results of the regression analysis together with model performance metrics and interactive graphs. 8. Applying the established statistical univariate and/or multivariate profiles to the online process variables is accomplished by comparing them with a golden track represented by the statistical univariate and/or multivariate profiles or by keeping the monitored current response between given statistical thresholds. 9. If the at least one monitored online process variable reaches a critical value such as ±2σ or ±3σ, the computer issues a warning signal via the digital instrument panel and/or the computer never allows the monitored online process parameters of the base components to exceed the critical values during a process by applying new set points to the established statistical univariate and/or multivariate profiles. 10. The machine learning model represents an accurate model of the production base and is used to drive a production process via real-time model predictive control (MPC), wherein the parameters of the MPC are mapped to the base components to monitor the online process variables, and wherein correction of the detected anomalies, interruptions and/or failures to improve the production process is accomplished by generating optimized set points and applying them to the established statistical univariate and/or multivariate profiles. 11. The dashboard indicates the performance of a product via the online process variables and predicts the performance of a batch currently being produced or recently produced via historical data of the product by analyzing a given quantitative input. 12. The online process variables are collected by the computer of the respective base process control system and sent to the respective computer operating the machine learning model after using the queries and/or available production process metadata from the base. 13. For statistical multivariate profiling, the data of each batch is reduced to at least one potential variable, and then the at least one potential variable is monitored. 14. To find similarities between programs in univariate, bivariate, and multivariate spaces, mathematical calculations such as Euclidean space can be used to sort all programs with the same topology from most similar to least similar programs for a particular program.

One further solution to the task provided is a system for operating a production base, which includes the following components of all existing hardware base equipment: a base process control program, a machine learning model executed by a software, a GUI A digital dashboard in the form of a database, a computer for operating the software, the digital dashboard, and a base process control unit for operating the base process control program, wherein all system components are connected via a closed circuit network The circuit structure is connected and the system is configured to perform the previously disclosed method. The system can then be set up to keep the ongoing process between these tolerance limits and send new set points back to the base. The system further enables the user to upload a model file and connect it to the program or batch(s). The uploaded model may be used to generate the univariate and multivariate contours or golden trajectories and the interpretation plots for use in a program of the outputs from the model-based analysis.

Preferred further developments of the system of the present invention include (for example but not limited to): Each component communicates with the base through the closed network using an OPC-UA protocol, where each process variable name indicates its respective component by a first code, which enables the system to assign each process variable to its specific component and It is monitored by its address in the available production program metadata. The digital dashboard enables the user to manually assign component names via a given function.

Further solutions include a computer program product and a non-transitory computer readable medium, with the computer program product stored thereon, including instructions for self-service advanced analysis when operating a manufacturing site as previously described. When executed on one or more processors of a system, these instructions cause the system to perform the methods previously described.

The method and system of the present invention can be applied to any suitable production base. Figure 1 shows in a schematic way the components of this preferred embodiment of such a system 1 using its assigned modular production plant 7 as a base. In this preferred embodiment, production-related data 2 created by executing production processes in the modular production factory 7 is collected and stored, such as sensor data, raw material data, process data and/or from factory machines. Quality information. The production plant 7 is preferably controlled by at least one control unit 5 using one or more specific control programs 6 to control the components of the modular production plant 7 and to execute the production program(s). The control unit or computer 5 is preferably located directly in the modular production plant 7, but a remote control unit 5 is also feasible, as long as a suitable data connection between the control unit 5 and the components of the plant 7 is ensured, such as some kind of network road. To perform the method of the present invention, a software 8 in the form of a SW analysis tool 8 has been created which uses different AI algorithms 3 in the form of a data-driven model 3 to analyze production-related data 2 as stored historical data or As "real-time data" collected in real time through control program 6. Which AI algorithm or model type 3 (e.g. XGBoost, decision tree, random forest, etc.) will be used depends on the specific use case and the method of the present invention generally does not depend on a specific model type. The method itself and its respective method steps are shown in Figure 3, wherein Figure 2 gives a schematic overview of the previously described state-of-the-art approach to facilitate comparison with the method of the invention. The software 8 further operates on a suitable analysis computer 4, which may be the same as the local control unit/computer 5 of the modular production plant 7, but preferably a separate computer 4 is used. The software 8 further provides a graphical user interface (GUI) in the form of a digital dashboard 10 on a screen, which enables a user 9 to operate the software 8 and use specific analysis functions of the software 8 .

Since this advanced SW analysis tool 8 now combines many of these functions and features, this tool 8 can be used to address different use cases leading to many different preferred embodiments, two of which are described in the following sections. Use case 1 – Time series analysis and real-time monitoring/control of factories

The SW analysis tool 8 provides a dashboard 10 via a GUI that displays possible analysis functions/features that can be selected by the user. Figure 4 shows the main menu of the GUI 10, which enables the user 9 to perform a process monitoring of the factory 7. For the first described embodiment, the use case is a multi-variable monitoring implemented by performing the following steps: 1-) A user 9 (e.g. an operator of the factory 7) wants to drive a process and monitor the process. 2-) Each process step (a factory will have several such steps) has dozens if not hundreds of variables. Therefore, the operator 9 selects the good batch from the drop-down menu of the relevant project in the dashboard 10. 3-) The dashboard 10 collects data for the selected batch. The calibration data is used to build a suitable data-driven model 3 and cross-validate it. This can be done for several suitable model types. After hyperparameter tuning, the best model 3 is selected as the final model 3. Based on this data-driven model 3, the dashboard 10 generates a so-called multivariate golden track. 4-) Next to the golden track, there will also be statistical variables such as the standard deviation of the golden track. In this case, three standard deviations are used as a critical value. 5-) Once the model and golden track are ready, any process data can be selected for comparison to identify any outliers, deviations and drifts. Clicking a timestamp from any time series plot (such as Figures 4 and 5) will generate a new plot. This newly generated graph (Figure 10) shows three different bar graphs of each parameter at the selected timestamp: the average of the golden process, the value of the selected process, and the percentage change of the selected process parameter relative to the average process parameter at the selected time. 6-) After the golden track is ready, the user 9 starts the process and starts monitoring. 7-) The dashboard 10 uses OPC-UA or other protocols based on the selected project to reach the laboratory equipment of this project. 8-) The dashboard 10 uses multiple threads and queries the real-time value from each piece of equipment approximately every 5 seconds. This time span can be set and must be done in advance by the user 9 through the GUI. 9-) The collected real-time values will be bundled and finally, the generated model 3 from the selected good batch will be used to transform and generate real-time multivariate scores. 10-) Then, it is used as an early warning system or to identify interruptions based on a multivariate profile. 11-) The user 9 can now set/create an alarm. If the multivariate profile of the ongoing process reaches a calculated limit, for example, usually 2σ or 3σ; the user 9 will get an alarm or a notification and can stop the production process. In another preferred embodiment, this last step can also be done by the tool 8 itself, thereby making the process/driver fully automated at the same time. Figure 5 shows the results from the tool 8, presented via the GUI 10 and the dashboard 10 respectively. 12-) Instead of generating a model 3 in step 3-), in a further preferred embodiment, an external model 3 can also be uploaded to the dashboard to generate such golden tracks or used to drive the process completely autonomously, for example using a feedback controller. Use Case 2 – Batch Level Analysis

This preferred embodiment covers a different use case. Here, a specific batch will be analyzed by performing the following steps: 1-) The user/operator 9 wishes to perform an analysis to optimize future processes or to identify any historical process dynamics/issues in the process, final/intermediate quality or raw materials. 2-) Next, the user 9 selects the program to be analyzed. If the selected project is or has been completed externally, this production data can be uploaded and edited interactively in the dashboard 10 2 . 3-) Analysis methods (such as statistical methods, machine learning or deep learning) are predefined or selectable. 4-) The user 9 is also given the option to select one of the calibrators. These calibration methods are predefined for each machine learning method. But it can also be a voluntary choice. Hyperparameter optimization methods and cross-validation techniques can also be configured voluntarily. 5-) Finally, the user 9 clicks the "training" button. Dashboard 10 collects the data of the selected batch. Calibrate the data, build a data-driven model 3, and perform cross-validation on it. Figures 6 and 7 show the cross-validation procedures of the respective training and test data sets of an XGBoost algorithm. Again, after hyperparameter tuning, the best model 3 will be selected as the final model 3. This final model 3 will be interpreted using some functions such as "permutation feature importance", "SHAP" or "Gini importance", "Hotelling's T2 distribution" or "to model's distance". Figure 8 shows the results of the permutation importance of the XGBoost algorithm. On the other hand, Figure 9 shows Hartling’s T2 distribution. Here, shown via the 95% confidence ellipse in the data output: some procedures are outside the ellipse. Here, any control chart can be used with the model and such charts can be program dependent. 6-) Once model 3 is ready; user 9 can control the interpretation diagram. Based on these diagrams, the most basic characteristics of the program can be identified and optimized. Once entity dependencies are understood from these graphs, this results in an overall improved program output. 7-) Similarities can be drawn for a specific program or for all programs. This can be achieved, for example, using methods based on distances between programs in a multivariate space. This reduces final quality variation by selecting the most similar materials. Use Case 3 – Change Analysis

This embodiment covers a variation analysis of a process. This variation can have multiple purposes. Since processes mainly consist of multiple stages, such single analyses can be performed in a more structured manner rather than performing multiple single analyses. This variation analysis can be used to detect in a single variable or multivariate manner whether any type of parameters and measurements are out of control or vary during a process for any reason, which may or may not have an undesirable impact on the overall process quality or the final product. 1-) Unless there is no connection to the factory/module or any storage, the user must upload a data set in the application. If a data set containing all steps and process types is uploaded, using "prefix", "last code" or any regular expression/pattern, the steps of a process will be analyzed. If there is a direct connection to the factory/module or any storage, the steps of a process will be collected along with the data. 2-) The user selects the process/dataset/batch. 3-) The application identifies patterns/steps in the process/dataset/batch and performs multiple batch-level and time-series analyses of the identified steps in the process using convenient machine learning algorithms (such as PCA/PLS). These steps/patterns can be defined based on a process. 4-) The results will be plotted using selected σ bounds (but can be defined as 2 and 3σ) or confidence intervals. 5-) The results will be investigated in multivariate and univariate space using a system of rules, such as violations of Nelson's rule: a) One or more points outside 3σ of the mean or a certain confidence interval b) Nine or more points in a row on the same side of the mean c) Six or more points in a row increase or decrease continuously d) Fourteen or more points in a row increase and then decrease in an alternating manner e) Two of three points in a row are outside 2σ standard deviations f) Four or more of five points are more than one σ standard deviation away from the mean in the same direction as the mean g) Fifteen points in a row are within 1σ standard deviation h) Eight points in a row are not within 1σ and the points are in both directions Some steps of these investigations can be automated. 6-) If these results indicate no violation of the defined/selected rules, the application will create a certificate that proves that the variation within the production process of the selected batch is within a specified tolerance.

1: System for self-service advanced analysis 2: Production-related data 3: Model/AI algorithm 4: Analysis computer 5: Factory/base control unit/factory/base control computer 6: Factory/base control program 7: Modular production factory/base 8: SW analysis tool/software 9: User/operator 10: Digital instrument panel

The methods, systems and software products according to the present invention and their functionally advantageous developments will be described in more detail below using at least one preferred exemplary embodiment with reference to the associated drawings. In the drawings, corresponding elements have the same element symbol.

Graphic display: Figure 1: A production base using the system of the present invention to enhance production efficiency of the base Figure 2: An overview of the latest technology production data analysis Figure 3: An overview of the production data analysis of the present invention Figure 4: Main menu of the GUI used to control the analysis software Figure 5: Results of a multivariate analysis via GUI Figure 6: The first part of the cross-validation results of a batch-level analysis Figure 7: The second part of the cross-validation results of a batch-level analysis Figure 8: Interpretation of the final model using a ranked feature importance Figure 9: Final batch quality assessment Figure 10: Comparison of averaged features from the best program with a selected program

1: A system for self-service advanced analysis

2:Production related information

3: Model/AI algorithm

4: Analyze the computer

5: Factory/base control unit/factory/base control computer

6: Factory/base control program

7: Modular production plant/base

8: SW analysis tools/software

9:User/Operator

10: Digital instrument panel

Claims (20)

A method for improving a biological/chemical production process executed by at least one computer (4, 5) on a production site (7), which has the following steps, including: Collect and prepare data (2) about the production process in the form of program variables by the at least one computer (4, 5); Optimizing and calibrating a suitable machine learning model (3) by the at least one computer (4, 5) depending on the program variables; A statistical univariate and/or multivariate profile of a unit operation of the production site (7) is established by the at least one computer (4, 5) via the optimized machine learning model (3); provided by the at least one computer (4, 5) by applying the established statistical univariate and/or multivariate profiles to at least one online process variable to detect anomalies, interruptions and/or failures in the production process A digital dashboard (10) monitors at least one online process variable of the production process; and Adapt the production process by correcting the detected anomalies, interruptions and/or malfunctions and/or instructing the user (9) to interrupt the production process if required quality standards are violated. The method of claim 1, wherein the machine learning model (3) is established by analyzing the process variables via a first algorithm, wherein the analysis result is displayed to a user (9) via the digital instrument panel (10), and then the user (9) selects a suitable model type (3) depending on the analysis result, and then the selected model (3) is established and optimized by the at least one computer (4, 5) by applying a suitable machine learning method such as principal least squares regression, XGBoost or random forest regression. The method of claim 2, wherein the user (9) selects at least one batch as the source of the online process variables for the production process and also selects specific process variables to be monitored via the digital instrument panel (10). Such as the method of request item 3, where For statistical univariate profiles, at least two different batches and at least one program variable are selected to calculate statistical characteristics such as mean, ±2σ, and ±3σ. A method as claimed in claim 1, wherein an existing suitable machine learning model (3) is used by the first algorithm by analyzing the process variables, wherein depending on the analysis results, a suitable model (3) is imported and optimized, which can completely autonomously drive the process on a given optimization golden track without any further user input. For example, the method of request items 2 to 5, among which This first algorithm uses several interpretive functions, such as permutation feature importance, Shapley Additivity Interpretation (SHAP) or Gini importance, to find and analyze basic program variables. A method as claimed in claim 1, wherein the machine learning model (3) further comprises a physical model resulting in a hybrid model. Such as the method of request item 1, where The machine learning model (3) uses specific secondary algorithms, such as principal component analysis (PCA), partial least squares (PLS), random forest, extreme gradient boosting (XGBoost), support vector machine, Gaussian program, multi-layer perceptron, They may be linear, nonlinear, or a neural network machine learning method, and use each second algorithm to complete a regression analysis and provide the results of the regression analysis along with model performance metrics and interactive charts. Such as the method of request item 1, where The created statistical univariate and/or multivariate profiles are applied to the online process variables by comparing them to a golden trajectory represented by the statistical univariate or multivariate profiles or by comparing the monitored The current reaction is completed while remaining between given statistical thresholds. Such as the method of request item 1, where If the at least one monitored online program variable reaches a threshold value such as ±2σ or ±3σ, the at least one computer (4, 5) issues a warning signal through the digital dashboard (10) and/or the computer ( 4) Never cause the monitored online process parameters of the base components to exceed the thresholds during a process by applying new set points to the created statistical univariate and/or multivariate profiles. value. The method of claim 10, wherein the machine learning model (3) represents an accurate model of the production base (7) and is used to drive a production process via real-time model predictive control (MPC), wherein the parameters of the MPC are mapped to the base components to monitor the online process variables, and wherein the correction of the detected anomalies, interruptions and/or failures to improve the production process is accomplished by generating optimized set points and applying them to the established statistical univariate and/or multivariate profiles. The method of claim 1, wherein the dashboard (10) indicates the performance of a product via the online process variables and predicts the performance of a batch currently being produced or recently produced by analyzing a given quantitative input via historical data of the product. The method of claim 1, wherein the base program control unit (5) operating the machine learning model (3) and the analysis computer (4) are separate computers and the online program variables are collected by the base program control unit (5), and the base program control unit (5) sends the variables to the analysis computer (4) after using each query and/or available production program metadata from the base (7). Such as the method of request item 1, where For statistical multivariate contouring, the data for each batch is reduced to at least one latent variable, and then the at least one latent variable is monitored. Such as the method of request item 1, where To discover similarities between programs in single, bivariate, and multivariable spaces, mathematical calculations such as Euclidean space can be used to classify all programs with the same topology from the most similar and least similar programs for a particular program. Sort. A system for operating a production base (7), which includes the following components of all existing hardware base equipment: a base program control program (6), a machine learning model (3) executed by a software (8), a digital instrument board (10) in the form of a GUI, a database, a computer (4) operating the software (8) and the digital instrument board (10), and a base program control unit (5) for operating the base program control program (6), wherein all system components are connected via a closed network structure and the system (1) is configured to execute the methods as claimed in claims 1 to 14. A system as claimed in claim 16, wherein each component communicates with the production base (7) via the closed network using an OPC-UA protocol, wherein each process variable name indicates its respective component by a prefix, which enables the system (1) to assign each process variable to its specific component and monitor it via its address in the available production process metadata. A system as claimed in claim 17, wherein the digital instrument panel (10) enables the user (9) to manually assign the component names via a given function. A computer program product including instructions that, when executed on one or more processors of a system (1) operating a production site (7) as claimed in claim 16, cause the system (1) to execute: A method for requesting any one of items 1 to 15. A non-transitory computer-readable medium having computer-executable instructions stored thereon which, when executed on one or more processors of a system (1) as claimed in claim 16, cause The system (1) performs the methods of claims 1 to 15.