Method for monitoring and regulating a technical plant process, and technical plant

Abstract

The invention relates to a method for monitoring and regulating a technical plant process, in particular an industrial manufacturing process for producing a product, in which - a profile representing the variation over time of at least one process parameter (P) such as pressure, temperature, geometric and / or optical features and energy consumption is recorded, - a momentary state of the process (Zi) is continuously determined on the basis of the at least one process parameter (P) such that a profile of states of the process (Zi) is obtained, wherein a respective state of the process (Zi) is a an indicator of the quality of the plant process or of a product produced by the plant process, - a trained model (M), which includes a probability matrix (W) which indicates probabilities (Wij) for a change of state (Zij) between different states of the process (Z), is used for the plant process, - a transition probability (Wij) for the change of state of the momentary state of the process ( Zi) into another state of the process (Zj) is determined on the basis of the model (M) and - a measure for achieving a predefined target variable (X) when the plant process is being carried out is initiated on the basis of the momentary state of the process (Zi) and the determined transition probability (Wij).

Description

[0001] FDST Patent Attorneys, Nuremberg Page 1

[0002] P230842P2-MD / JB

[0003] Description

[0004] Methods for monitoring and controlling a technical plant process and technical plant

[0005] The invention relates to a method for monitoring, controlling, and regulating a technical plant process, in particular an industrial manufacturing process for producing a product. The invention further relates to a technical plant for carrying out such a technical plant process.

[0006] When manufacturing a product, for example using a continuous manufacturing process such as extrusion or a discontinuous manufacturing process such as injection molding, a multitude of process parameters come into play that reflect and / or influence the quality of the manufactured product. These process parameters are often subject to fluctuations, which are reflected in the product quality. To avoid poor quality and scrap, continuous monitoring of the production process is therefore usually implemented. Due to the complex interrelationships involved, such monitoring is often time-consuming.

[0007] In the present context, a discontinuous manufacturing process is generally understood to mean a process in which several work steps are carried out repeatedly and, in particular, immediately one after the other within a cycle, whereby this manufacturing process can be treated as a quasi-continuous process due to the repeated execution of the cycle.

[0008] When the term "plant" is used here, it is understood in particular to mean a plant system, as an interconnected network of several technical plants and machines that work together to control a complex process.

[0009] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 2

[0010] Based on this, the invention aims to provide a method for monitoring, controlling and regulating a technical plant process, as well as a technical plant for carrying out such a plant process, in which reliable monitoring and control and thus real-time optimization of the plant process takes place.

[0011] The problem is solved according to the invention by a method for monitoring, controlling, and regulating a technical plant process with the features of claim 1, and by a technical plant for carrying out such a technical plant process with the features of claim 27. The advantages and preferred embodiments mentioned with regard to the method are also transferable to the plant mutatis mutandis. For carrying out the plant process, the plant includes, in particular, a plant control system that is suitable for executing the various steps of the method during operation.

[0012] The plant process is primarily an industrial manufacturing process for producing a product, but it is not limited to this. For example, the steps described below can also be applied to the operation of technical plants that do not serve to manufacture a product, such as plants for providing heat or cold (heating systems).

[0013] The following steps are carried out in the procedure:

[0014] A time-dependent profile of several process parameters, including but not limited to pressure, temperature, geometric and optical characteristics, and energy consumption, is continuously recorded. Process parameters also include derived quantities such as mean, median, minimum, maximum, slope, curvature, range, and variance of such (primary) process parameters. These derived quantities are also referred to here as secondary process parameters or covariables.

[0015] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 3

[0016] Data acquisition encompasses both the direct measurement of (main) process parameters, for example using sensors, and their computational determination and derivation from other quantities. Process parameter values ​​can also be determined through image acquisition and corresponding image analysis. For each process parameter, current actual values ​​are preferably recorded and used, as well as target values, either alternatively or additionally.

[0017] In particular, the plant's control system contains at least one learned model for the plant process, also known as a state model. This is specifically a hidden Markov model. This model contains at least one, and preferably several, probability matrices, which are preferably derived from historical data. A probability matrix is ​​designed as a transition probability matrix, indicating the probabilities of a change between process states. The probability matrix thus reveals the probability that the plant process will transition from one current process state to another. This transition between two process states is referred to here as a state change.

[0018] Furthermore, a probability matrix preferably also contains probabilities for a so-called emission (hereinafter also referred to as emission probability), with which a specific process state generates at least one observable process parameter and preferably several process parameters. That is, the probability matrix also contains a statement about the probability with which a specific value of a particular process parameter is present in the respective process state. Typically, a probability distribution of the respective process parameter is assigned to each process state, indicating the probability with which which value of the process parameter is emitted by the respective process state.

[0019] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 4

[0020] A key aspect and an important basic assumption for using the state model is that hidden, not directly perceivable process states essentially determine and generate the measurable / observable process parameters at any given time. In other words, the measured process parameters allow inferences to be drawn about the hidden process states.

[0021] Based on at least one process parameter and preferably on several process parameters, a current process state is continuously determined, so that a progression of process states is obtained, wherein each process state defines at least one indicator for the quality of the plant process or of a product manufactured with the plant process.

[0022] In this context, "continuous" means that a process state is determined at regular intervals, for example, less than one minute, specifically between 5 and 30 seconds. In production plants and industrial processes, these intervals are often less than one second. In other systems, such as those used for heating or cooling, the intervals are frequently longer, for example, up to one hour, or between 10 and 30 minutes. Each process parameter, which varies during the process, is therefore assigned a corresponding process state, resulting in a sequence of process states throughout the process.

[0023] This process is also referred to as decoding. In particular, it involves classifying and thus assigning the various fluctuating values ​​of the process parameters to the different process states based on the stored and trained model (state model). Therefore, decoding determines the most probable underlying sequence of process states.

[0024] Each process state is generally assigned and allocated at least one quality parameter, which in particular allows conclusions to be drawn about the quality of the plant process or the manufactured product, or about the quality or grade of a

[0025] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 5 The process states determined in each case are preferably distinguished and described by characteristics and parameters of a large number or all process-relevant parameters or secondary process parameters. This comprehensive description, hereinafter also referred to as state description, of the individual process states is particularly important for the desired optimization of the plant process. For example, one process state represents the highest product quality, another medium product quality, and yet another minimal energy consumption (for example, with medium product quality), etc.It is also possible that the quality parameters could serve as a termination criterion, for example, if a specific process state leads to a defective product. In such a case, the plant process is interrupted or terminated to prevent the production of defective products. Considering these process interruption or termination states / final states is particularly necessary to ensure an overall reliable and therefore sustainable process flow.

[0026] It should be emphasized that, based on a process state, a large number (e.g. more than 5) of process parameters (with predefined value ranges for the respective process state) are assigned to it, which characterize this process state.

[0027] To determine the current process state, preferably only a subset of the process parameters assigned via the process state are used, and in the simplest case, only a single process parameter. After the process state has been determined, the various process parameters with defined value ranges are therefore linked to this process state via the state description.

[0028] Furthermore, based on the current process state and the determined transition probability, a measure is initiated to achieve at least one predetermined target value during the execution of the plant process.

[0029] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 6

[0030] Preferably, several target variables are used as predefined target variables, or sometimes only a single target variable is used as a predefined target variable.

[0031] Preferably, the end of production, an interruption, and / or a respective process termination—for example, due to a machine or process error—also constitute an (observable) process state. In a preferred embodiment, such a process state is also assigned to the model; that is, such (observable) process states are also represented by the model.

[0032] In this context, "process termination" refers to any interruption or critical disruption of the technical plant process, including process errors, that lead to instabilities and, in particular, to complete process failures or shutdowns. Preferably, these different types of process terminations—namely, instability and, in particular, process failure and shutdown—are distinguished. The model preferably also models and considers such process states and, in particular, differentiates between them.

[0033] Furthermore, the model preferably describes the entire plant process, thus mapping it, for example, like a digital twin using an algorithm. This means the model mathematically describes the plant process and makes predictions, for example, about how the process parameters behave and develop over the course of the plant process.

[0034] The method therefore begins by structuring and simplifying the plant process through classification into different process states. Simultaneously, by checking the emission and transition probabilities for a state change, potential changes in the process states are identified early on, allowing countermeasures to be initiated in a timely manner. Emission probability is, in particular, the probability with which a respective process state is emitted. Transition probabilities can also be used to determine, for example, whether a transition to a lower-quality state will occur.

[0035] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 7 is expected and therefore no action is required to achieve the desired target size.

[0036] The measure in question is, for example, an automated intervention in the plant control system to influence at least one of the process parameters. Alternatively, the measure may simply involve the output of information such as a warning or a recommendation for action. For this purpose, an assistance system is implemented that provides recommendations for action to the plant operator, i.e., the person operating the plant. Specifically, the assistance system includes a monitor on which the recommendations for action are visualized or displayed. In addition, the entire plant process is preferably also visualized, in particular the current process state and / or the progression of process states, thus enabling a more concise and understandable visualization.

[0037] In this context, the term "target variable" generally refers to a value of at least one of the process parameters, or a derived secondary process parameter, which is the desired target and should be achieved. The procedure thus continuously checks whether the current process state is suitable for achieving the target variable.

[0038] In a preferred embodiment, a final state of the process is considered as a process state. Preferably, a termination criterion of the plant process is used as such a final state and as a further process state. That is, the model also includes a final state of the process. The final state characterizes the end of the process. The final state can be of different natures. For example, the final state is a termination and / or interruption state resulting from a machine failure, a product failure, or a process failure, if, for example, one of the process parameters, such as temperature, or a secondary process parameter derived from this (main) process parameter, such as the variance, leaves a permissible range of values, in particular a tolerance range. Alternatively, the final state is the end of production if...

[0039] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 8 This means that no disruption has occurred and, for example, a predetermined production quantity has been reached or a change in order has taken place, etc. By implementing the respective (observable) end states as process states to be considered, the process capability of the process is mapped, among other things. Based on the transition probability to a critical end state, such as machine failure or process failure due to a critical parameter deviation, it is possible to derive the probability with which a process will transition to or terminate in this state. This also allows, for example, timely alarms to be triggered if a process is in or approaching a process state that involves a transition to one of the critical end states (e.g., machine failure).The probability of reaching this final state increases.

[0040] The production process, especially in the manufacturing of a product, is typically subject to fluctuations due to varying external factors. While the fundamentally same production process is still being used, for example, to manufacture the same product, the process unfolds differently due to these varying external influences, resulting in different process variants or potentially occurring in different variations.

[0041] In a preferred configuration, the model contains several different model variants, each corresponding to a different process variant. The model therefore has multiple model variants. This measure, along with the assignment to a specific model variant, allows for a more accurate representation of the actual, current plant process within a single model. This assignment to a suitable model variant simultaneously enables monitoring, control, and optimization of the actual plant process.

[0042] The model variants are used in particular for process control (control / regulation of the plant or the technical plant process), and for prioritizing...

[0043] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 9

[0044] Interventions for plant control, the associated selection of measures and determination of the intervention times were used.

[0045] Preferably, several or even all of the stored model variants are compared with the current plant process during operation. The model variant that best represents the current plant process is then preferably selected as the model for determining the process states and their transition probabilities. The plant process is therefore observed for a certain period and continuously compared with the different model variants. The model variant that best matches the actual course of the plant process is then preferably selected and used for the previously described method of monitoring the plant process.

[0046] Preferably, comparisons with other model variants are continued throughout the process. If it is determined that another model variant better represents the current plant process, a switch to this other model variant is preferably made automatically. The model variant is determined using the highest / best degree of agreement or quality measure, which is also referred to as the adjustment measure. Switching between model variants accounts for changing influences on the process, as these often do not remain constant. For example, temperature increases, increasing wear, and changes in the input batch of material affect the process. Therefore, switching between model variants during operation ensures a realistic representation of the process with changing influences.

[0047] The adjustment measure therefore serves as a criterion for how well the current plant process is represented by a given model variant. Several differently determined adjustment measures can be used.

[0048] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 10

[0049] A first adjustment measure is, for example, a measure determined by the likelihood function, which is known per se, in particular a measure determined by the log likelihood function.

[0050] Another adjustment measure is, for example, a difference measure. The difference measure generally indicates how far apart the values ​​of a process parameter are in two comparable processes. These processes could be, for example, the current actual process and a process simulated with the respective model variant. To calculate the difference measure, the difference in the values ​​of a process parameter between the two processes is determined—especially at comparable points in time—and the difference measure is derived from these individual differences at the respective points in time, for example, by summing them. For instance, the individual absolute differences or their squares are summed or averaged. To do this, the two processes are preferably shifted relative to each other on the time axis in such a way that the difference measure has a minimum.

[0051] The adjustment measure is time-dependent, meaning it changes during the ongoing operation of the plant process and / or depending on a specific time window. In other words, the adjustment measure changes over the course of the plant process and, in particular, over the course of at least one process parameter.

[0052] If, for example, the plant process is considered over different time windows, different adjustment measures typically result, whereby this variation is preferably not solely due to the length of the time window, but also results in particular from the process dynamics contained therein.

[0053] Preferably, the same time window is used to determine the adjustment measure for both the simulated course of the plant process and the current actual course of the plant process.

[0054] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 11

[0055] Due to the time dependency of the adjustment measure, particularly its dependency on the selected time window and the process progression within it, a continuously updated adjustment measure emerges over the course of the plant process, and especially over the course of at least one process parameter. The time window is continuous and therefore shifts as the process progresses. This time window is preferably variable, meaning it can, in particular, increase in size and expand. Alternatively, it can remain constant or decrease in size. The continuously updated adjustment measure can therefore also be considered a sliding adjustment measure.

[0056] In a preferred embodiment, the determination (and application) of the (currently best) model variant takes place in real time during the ongoing plant process.

[0057] In particular, adjustment measures for the different model variants are continuously determined (permanently or at fixed times, e.g., in the range of seconds) and compared with each other.

[0058] The system preferably switches between model variants during the ongoing plant process. This occurs automatically, without requiring operator intervention. Alternatively, switching can also occur only after operator approval. The model variants are thus continuously, preferably automatically and in real time, compared with the current plant process, and switching between model variants can occur during the ongoing process depending on the determined adjustment parameters, or it can occur automatically.

[0059] Established methods are used to determine the degree of agreement and thus the degree of fit. For example, the behavior of one or more process parameters of the actual plant process is compared with a multitude of simulated parameters from the various model variants.

[0060] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 12

[0061] Plant process flows are compared. The values ​​from the simulated plant processes are preferably represented in distributions of the fit measure for the respective model variant. For this purpose, the previously described likelihood function or the described difference measure, for example, are used as fit measures.

[0062] In a preferred embodiment, a distribution of at least one adjustment measure is determined for each model variant, and these distributions are taken into account when selecting the respective model variant (as the prioritized model). The distribution is generated, in particular, using simulations. In these simulations, a plant process is repeatedly simulated with each model variant, especially for a predefined time period (time window). This results in a large number of simulated process trajectories (with each model variant), for example, at least 100 or at least 500. These different simulated process trajectories are compared, and a large number of adjustment measures (e.g., as a difference measure) are obtained. Each adjustment measure, especially the difference measure, indicates a value for the agreement between two simulated process trajectories. Alternatively, orAdditionally, the fit measure, in particular the likelihood function, provides a value for the degree of agreement between the respective simulated process flow and the model. Preferably, each simulated process flow is compared with every other simulated process flow. The numerous individual fit measures determined in this way are also referred to as simulated fit measures. They exhibit a distribution. This distribution forms the basis for determining the model variant to be prioritized.

[0063] The distribution of at least one adjustment measure that is considered for a given model variant is therefore - as described - a distribution of adjustment measures for a given model variant determined by simulations.

[0064] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 13

[0065] This distribution indicates the probability of obtaining a specific fit measure, i.e., the probability of assigning this fit measure to the respective model variant. This probability is preferably used to select a model variant as the prioritized model.

[0066] To select the model variant as the prioritized model, these model variant adjustment measures (the simulated adjustment measures), i.e., the previously described distribution, are compared with the actual resulting values ​​of the respective adjustment measure obtained for each model variant in the (current) investment process. In other words, for each model variant, an adjustment measure is determined between the trajectories simulated with that model variant (currently) and the actual trajectory. This adjustment measure is also referred to as the current (actual) adjustment measure.

[0067] In particular, the model variant is selected where the current adjustment measure has the highest probability of assignment.

[0068] As already explained, the temporal progression of the adaptation measure of the plant process is continuously determined and considered and compared with the distributions of the adaptation measures of the different model variants at identical time windows.

[0069] Due to the continuous progress of the plant process, the distributions of the adjustment measures change continuously, which means that the distributions of the adjustment measures (for the respective model variant) must be calculated as a function of time relative to the actual plant process under consideration, in order to ensure a regular optimal determination of the possibility of assigning the actual plant process to the different model variants, especially in real time.

[0070] The adjustment measure is generally time-dependent, meaning it varies over time.

[0071] Preferably, it is a continuously updated adjustment measure. With

[0072] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 14 As the process progresses, the adjustment measure is always determined for, e.g., a given running or expanding time window.

[0073] As explained previously, the selection of the model variant is based on a distribution of the fit measure determined through simulations. Due to its time-dependent nature, this distribution is also time-dependent.

[0074] Preferably, the distribution is determined continuously throughout the process. This means that during the actual process, the time-dependent distributions are determined based on simulations for the relevant time window of the plant process. Alternatively, these distributions can be determined beforehand, i.e., not during the process.

[0075] The selection of the appropriate model variant is therefore preferably not based solely on individual (current, actual) fit measures of the various model variants, whose values, for example, represent an optimum for the current process. Rather, the selection is based on the time-dependent distributions of the (simulated) fit measures that show the greatest agreement with the (actual, current) fit measures of the current process determined over time, preferably with the highest probability of assignment, as previously explained.

[0076] The current, actual fit measure refers to the performance measure (i.e., adjustment measure, e.g., difference measure) of the respective model variant compared to the actual process flow. In contrast, the simulated fit measure refers to a performance measure between the process flows simulated with the model variant.

[0077] For example, if one of the process parameters exhibits a specific value or value profile during the actual process, the model variant with the best current, actual fit (e.g., with the smallest difference between the actual) will not be used as the basis for selecting the prioritized model variant.

[0078] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 15

[0079] The value and the value determined with the model variant are used. Rather, the model variant used as the prioritized model is the one in which the distribution resulting from the simulations yields the highest probability of assignment to the current, actual fit measure.

[0080] The time-dependent distributions of the adjustment measures of the different model variants can also be determined, for example, by Monte Carlo simulations, which depict possible trajectories of the model variants taking into account unknown environmental influences.

[0081] The time-dependent model assignment and switching offers the particular advantage that reliable model-based process monitoring and thus control of the plant process can be achieved even under varying, especially unknown, influencing conditions.

[0082] If none of the stored model variants adequately / validly reflects the current plant process, a new model variant is preferably established based on the actual plant process and assigned to the existing model variants.

[0083] In the preferred configuration, the model is generally continuously verified and adapted by the various plant processes during the course of these processes.

[0084] Preferably, the various model variants are classified into different classes based on characteristic features of the plant process in the respective process variant, with each class of specific model variants typically exhibiting certain characteristic similarities. For example, studies have shown that the different process variants differ characteristically in their behavior, depending on fluctuations in external influencing factors.

[0085] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 16

[0086] The process variants differed, for example, in their abruptness, i.e., in the frequency of state changes. Other process variants, in contrast, were characterized by a homogeneous progression and few state changes.

[0087] It has also been shown that the process variants differ, for example, in their pattern profiles, such as an increasing drift. In this case, the emission probabilities of the states might be skewed, or the transition matrix might tend to show higher transition probabilities to states that correlate with a higher trend value. Other variants, in contrast, are characterized by a homogeneous profile, in which the states are, for example, normally distributed and the state transitions occur less frequently or more uniformly between a few states.

[0088] Finally, historical analyses show that a class of process variants is characterized by at least one process parameter. Based on this classification, and especially in conjunction with the previously described assistance system provided to support the plant operator, it becomes immediately clear to the operator whether a critical plant process (process variant) is present, or whether the current plant process is, for example, particularly reliable. Particularly in the case of a process variant characterized by drift, especially intensive monitoring or intervention is required to ensure that the desired target value is not missed.

[0089] In a preferred embodiment, the individual characteristic classes of the specific model variants can contain further subclasses, which, for example, allows the monitoring of an ongoing plant process to be scaled (more finely).

[0090] In a suitable design, each class and, where applicable, its subclass is also assigned a quality indicator. This is displayed or made known to the plant operator, specifically via the aforementioned system.

[0091] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 17

[0092] Assistance system. This is achieved, for example, through color coding similar to a traffic light. In a preferred application, the operator can be advised of a targeted class change, including specific control measures. Similar to the change control for a process state, this allows for targeted influence on the entire characteristic process progression.

[0093] According to an alternative variant, the plant process is fully automated without the need for manual intervention.

[0094] For classification purposes, different (current) process flows or model variants are preferably segmented based on their historical or simulated trajectories and / or on model-specific key performance indicators. The different segments form the various classes.

[0095] In a preferred embodiment, similarity measures are calculated between the process flows or model variants, and these are grouped into clusters based on these similarity measures and thereby classified into classes. Model variants with a comparable similarity measure are thus assigned to a common class. The similarity measure is, in particular, the previously described fit measure. That is, process flows are classified based on their fit measure.

[0096] As explained previously, each class is preferably assigned at least one quality indicator, which allows for an assessment of the process or product quality.

[0097] The classification preferably allows conclusions to be drawn about various influencing factors, especially unknown environmental and / or material influences. In particular, the classification systematically categorizes previously unknown process influences. Their effects on the process are identified and preferably quantified.

[0098] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 18

[0099] The classification is preferably hierarchical, so that - as mentioned before - classes are further divided into subclasses, enabling a more detailed root cause analysis and process optimization.

[0100] The method described here, in its preferred further development, is characterized in particular by the fact that various determined process states already lie within a permissible tolerance range and yet differ from one another, and that, in order to achieve at least one predetermined target value, a measure is initiated—that is, control measures are taken on the plant process—even if a value for at least one process parameter, and more generally for a process-controlling parameter, remains within the permissible tolerance range. Thus, control measures are preferably taken even if there is not yet a risk of exceeding the permissible tolerance range. Typically, tolerance ranges are specified for certain process parameters, within which they are allowed to operate.The present method is characterized by the fact that the process states already differ within this permissible tolerance range, and measures are taken to optimize the values ​​within this range. This means that measures are preferably implemented even when the values ​​are recognizably still within the permissible tolerance range, i.e., when the model indicates that the values ​​are expected to remain within the tolerance range. This triggers optimization measures early on—for example, when the probability of the process transitioning to a more critical state increases (while the process parameter values ​​remain within the permissible tolerance range)—thus preventing defects in manufactured products or in the plant process.Secondly, optimization is enabled with respect to at least one predefined target variable that differs between the different process states, while at the same time ensuring compliance with the permissible tolerance limits.

[0101] Dies ist insbesondere auch das Resultat von mehreren Maßnahmen bei dem vorliegenden Verfahren. Unter Anderem beruht dies auch auf die nachfolgend noch

[0102] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 19 This document further explains the consideration and assignment of a process flow to, for example, quality data that is preferably linked to the process (see below for the explanations regarding "offline quality data"). Furthermore, this is additionally or alternatively based on the consideration of risk indicators, as explained below (see below for the explanations regarding the assignment of risk indicators to the process states or classes).

[0103] The target variable is optionally at least one predefined quality parameter relating to the plant process, such as temperature, pressure, or throughput time of process parameters, or a quality parameter of the product, such as a desired geometry that must be guaranteed. Alternatively or additionally, the target variable is energy consumption for the plant process or for the production of the product, and / or process speed. Preferably, the target variable can be selected by the plant operator. Another target variable is, for example, material consumption, which should be minimized.

[0104] These aforementioned target variables are often interdependent and mutually dependent. The available options therefore allow the plant process to be optimized with regard to a given objective. Of particular interest here, but not exclusively, are the target variables of minimizing energy or material consumption. Depending on the customer requirements of a manufactured product, for example, high quality may be required, or alternatively, only medium quality. This can mean, for instance, deliberately accepting a reduction in quality in the plant process to achieve medium quality while simultaneously optimizing energy or material consumption.

[0105] The desired target variable is selected based on external requirements, especially customer requirements. Alternatively or additionally, the target variable is selected based on...

[0106] - operational requirements

[0107] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 20

[0108] - based on current process or environmental conditions, and / or based on user preferences.

[0109] In a preferred configuration, several target variables are considered, for example, by offering them as options for selecting a given target variable. Alternatively, several target variables can be selected as predefined target variables to be achieved. In such a case, the plant process is therefore controlled with a view to achieving, or achieving as effectively as possible, several target variables.

[0110] The plant process is typically controlled with regard to a selected (or at least one) predefined target state; that is, the plant process is optimized specifically to achieve the target variable. Suitable measures, particularly intervention measures, and / or intervention times are derived and applied to achieve and stabilize the predefined target state. When controlling for multiple target variables, such as a quality parameter as the primary target variable while considering a secondary process parameter (also called a co-variable), such as energy consumption, the secondary process parameter is optimized, for example, within a predefined tolerance range for the quality parameter.

[0111] Preferably, the target variables are prioritized. This means that the target variables are weighted differently according to their importance. The predefined target variable, at least one, is selected based on this prioritization. Specifically, the target variable with the highest priority is chosen as the predefined target variable. If there are multiple predefined target variables, they are preferably prioritized.

[0112] The choice and prioritization of the target variable determines, in particular, whether, for example, quality improvement, energy saving, material usage optimization or throughput maximization is the main focus.

[0113] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 21

[0114] In preferred further development, the prioritization and, in particular, the selection of at least one predefined target variable is dynamic. This means that the prioritization changes during the course of the plant process. Specifically, the predefined target variable is also modified, so that the plant process is preferably aligned with the new target variable during the course of the process. The selection of the preferred process state (as the prioritized target variable) can therefore be changed dynamically during operation. Generally, this dynamic adjustment allows switching between target variables during operation.

[0115] The change (dynamic adjustment) of the target variable occurs, for example, on the basis of changed market requirements or customer requirements.

[0116] The selection, or at least a proposal, of a target variable, and thus in particular a preferred process state, is preferably automated. For this purpose, the various target variables are weighted with critical thresholds. If such critical thresholds are reached, the corresponding target variable is automatically reclassified with regard to its prioritization, e.g., its priority is increased or decreased.

[0117] In particular, target variables that signal a risk of quality loss or process interruptions by exceeding a critical threshold are preferably automatically (re)prioritized and preferably (upon reaching the critical threshold) linked to measures such as early warnings.

[0118] This dynamic prioritization achieves time-dependent target optimization. The weighting (prioritization) of the target variables adjusts over the course of the process, particularly depending on fluctuations in unknown environmental influences or material properties.

[0119] By optimizing the plant process with regard to the preferred process state defined by the target variable, for example, non-prioritized target variables are also preferably kept stable within defined tolerance ranges.

[0120] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 22

[0121] The method described here preferably provides information about the plant process, such as analysis results, state descriptions (including statistical information / properties), information about state changes, and / or predictions about the further course of the plant process, via a user interface, preferably in a suitable, structured form. Specifically, this structured provision is preferably used for further evaluations and / or visual representations, as well as to support plant operation, process optimization, or model optimization.

[0122] The user interface also allows for input and intervention in the plant process to control it in a targeted manner. This intervention takes place, for example, within the framework of the expert system described below.

[0123] This representation, and in particular the possibility of user input, is preferably generally provided by a suitable software program, such as an app, which is installed on a computing unit / user device.

[0124] The processing and provision of this information is preferably carried out by one or more computing units, which can be operated locally, centrally or in distributed systems.

[0125] In the following, "computing unit" refers to any physical or virtual device suitable for processing, storing, or controlling data and processes. This includes, but is not limited to, controllers, servers, processors, cloud systems, and database systems.

[0126] There can be multiple user interfaces or just one shared user interface. The user interface consists of standard input and output devices such as keyboards, screens, voice control systems, etc.

[0127] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 23

[0128] The user interface visualizes process-specific information, in particular the current process state (Zi), as well as preferably relevant trends, key performance indicators, model variant and / or class assignments, probabilities, and forecasts. Additionally or alternatively, cross-process information from multiple processes, machines, lines, or locations is visualized for comparison.

[0129] In the preferred variant, process-specific and / or cross-process simulations are used to evaluate the effects of different intervention strategies, specifically – in the case of cross-process simulations – both holistically and at the same time in a process-specific manner.

[0130] The interface preferably provides recommendations for action, prioritizations, and intervention options, particularly including the expected achievement of targets with regard to potential key performance indicators (such as quality, energy consumption, resource utilization, and throughput). These action options can be triggered by the user and / or initiated automatically. Ideally, the effectiveness of the selected action option is reported back and used, in particular, for further optimization.

[0131] Preferably, information is automatically exchanged between a process-specific and a cross-process analysis, so that both levels (process-specific and cross-process) are informed about specific and / or global measures, warnings and action periods, and coordinated intervention strategies are enabled.

[0132] The interface preferably includes a digital knowledge module that explains the correlations between process parameters, state changes, model variants and classes, and makes them usable for training purposes in particular.

[0133] In a preferred embodiment, the transition probability is determined by considering the previous course of transitions between preceding ones.

[0134] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 24

[0135] Process states are taken into account. This means that the model not only considers the current process state and determines transition probabilities for a state change based on the current state, but also incorporates the preceding history of state changes into the transition probability. As mentioned earlier, the model is trained on historical data. By comparing the historical trajectories of different state changes, a pattern or structure can be identified, revealing whether and to what extent a specific sequence of previous state changes influences a change in the current process state, allowing this prior data to be factored into the transition probability.To determine the current transition probability, the current process history is also considered; that is, it examines which process states existed immediately before the current process state and which influence the next state change. For example, only one previous process state (before the current process state) is considered, or two or more previous process states are taken into account.

[0136] This is taken into account in the model by means of a higher-order model, whereby a second-order model is understood to consider only the previous process state in addition to the current process state. A third-order model considers two previous process states, and so on.

[0137] Preferably, the (instantaneous) residence time in the current process state is considered when determining the transition probability. The model therefore preferably also includes a correlation between the residence time and the transition probability.

[0138] The transition probability is specifically represented as a function of the dwell time. That is, it is determined primarily as a function of the dwell time. This measure allows for the determination of an ideal / preferred

[0139] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 25

[0140] Determine the intervention time or period. This means that, depending on the functional relationship between the transition probability and the dwell time, an intervention time / period is derived, and an intervention measure is proposed or automatically executed within the intervention period. This allows for the prioritization of an intervention time, particularly in the case of so-called linked lines.

[0141] This correlation or dependency is preferably determined and derived from the historical trajectories of the various plant processes. For example, it can be observed that a particular process state is relatively stable and that the plant process remains in this existing state for an extended period. Alternatively, it can be observed that another process state is very unstable and tends to change after only a short time. The model therefore contains information about which process states develop stably and which become unstable over time.

[0142] The dependence of the transition probability on the residence time is represented in particular by a so-called hazard function.

[0143] This is determined, for example, using the statistical method of the failure rate (hazard rate). In the model, this is described in particular by a mathematical function (e.g., the hazard function).

[0144] Preferably, different hazard functions (increasing, decreasing, constant) are modeled. This allows, in particular, the determination of a time-dependent, adaptive transition probability.

[0145] A defined hazard function is also preferably used to assess the stability of the process state.

[0146] The transition probabilities are used, taking into account the duration of residence, particularly to determine suitable intervention times and / or to

[0147] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 26

[0148] Prioritization and optimization of competing target variables are considered. Appropriate intervention measures are initiated accordingly, if necessary.

[0149] Overall, this allows for the determination of a probability or current risk of a change in state. This probability is therefore time-specific, as it defines a timeframe for appropriate intervention.

[0150] The time-dependent transition probabilities (functions) are used, in particular, to determine a suitable intervention time for the plant process. Especially in multi-machine operation, this allows for a prioritization of intervention measures.

[0151] The transition probabilities contained in the probability matrix are therefore preferably not static values, but rather time-dependent values, where an (initial) transition probability is weighted by a time-dependent factor or function. These time-dependent transition probabilities allow us to determine the risk that the plant process is (currently) at risk of leaving its current state within a specific time interval. In other words, the transition probability varies with the duration of the process's current state.

[0152] The assignment of a given process state to at least one process parameter, and in particular to a set of process parameters, is preferably not static, i.e., not merely based on fixed, predefined value ranges for the at least one process parameter or for the various process parameters. In a preferred embodiment, the assignment is instead made taking into account an assignment probability stored in the model or determinable by the model. For example, for a given value range of a process parameter, the assignment to a specific process state is provided, whereby the probability of assignment to the process state for different values ​​of the

[0153] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 27

[0154] The probability of assigning a process parameter varies. For example, the probability for the mean value of this range is higher than for the extreme values. Simultaneously, the historical behavior of the process parameter and, preferably, the transition probabilities of previous process states are also considered in the assignment probability. This is particularly important in the case of a potential state change. By considering the assignment probability, the conditions for a state change, or for assignment to a new state, are smoothed and made variable, so that a state change is not necessarily or immediately assigned, for example, in the case of a short-term outlier in the value of a process parameter. This measure therefore stabilizes and smooths the overall behavior of state changes.

[0155] In a preferred embodiment, the transition probability for the transition from the first process state to the second process state is also considered when determining a change of state during the ongoing plant process, i.e., from a first, in particular the current / momentary, process state to a second, in particular a subsequent state, based on a change in at least one process parameter, and especially based on changes in several process parameters. Therefore, the transition probability for the transition between the two process states is additionally used when actually determining a change of state. That is, for the determination of a change of state, the assignment of the value of a process parameter to a process state is additionally weighted by the transition probability.If this probability is low, then a change of state will likely not be detected initially, even if the current value of the process parameter would allow for an assignment to the other process state. This also means that the progression of state changes is not based on hard, fixed criteria. Rather, by considering the transition probability, the progression of state changes is stabilized and smoothed, for example, in the case of short-term outliers in the values ​​of a process parameter.

[0156] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 28

[0157] In a suitable advanced training, the relevance of various process parameters for the state change between two process states is stored in the model. For assigning a parameter to a current process state, several process parameters are preferably used—that is, a complete set of process parameters (including any secondary process parameters). Within this set of process parameters, the relevance, as well as the (local) direction and / or strength of influence with regard to a state change, are stored for each parameter. The relevance, as well as the (local) direction and strength of influence, are in turn trained from historical data. The (local) direction of influence indicates in which direction the value of the respective process parameter must be changed to achieve a defined / desired state change.The degree of influence is a measure of how strongly the respective state is influenced by the respective process parameter and, in particular, how strongly a change of state is influenced by the process parameter.

[0158] Local influence direction / intensity refers to the current direction / intensity of influence of the respective process parameter, taking into account the other process parameters.

[0159] It is important to consider that each process state is characterized by a multitude of different process parameters, which, however, have varying degrees of relevance to that specific process state. Based on these parameters, it is therefore possible to determine precisely which process parameter needs to be changed and how, in order to bring about a targeted change of state or to stabilize the process state.

[0160] Identifying the relevance—that is, the significance and influence—of a particular process parameter with regard to a potential change of state enables the targeted manipulation of the plant process through specific measures. Specifically, the most relevant process parameter can be deliberately influenced, for example, to bring about a specific change of state or to stabilize a current state.

[0161] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 29

[0162] This information, e.g., about the relevance, direction of influence and / or strength of process parameters for the state change, is used in particular to

[0163] - To make changes of state explainable,

[0164] - To determine intervention measures and their prioritization and

[0165] - To selectively change process parameters in such a way as to achieve a desired target state or to avoid an undesired change of state.

[0166] The identification of relevance is preferably also deliberately used for targeted control of the plant process, which can be carried out manually or automatically. In a preferred embodiment, as part of an optimization measure for the plant process, at least one process parameter is specifically changed, optionally to stabilize the current process state (stabilization control) or to specifically bring about a different process state (switching control).

[0167] The selection and implementation of the optimization measure is based in particular on the model-based determined transition probabilities and preferably on the influence directions and strengths of the process parameters stored in the model.

[0168] At least one process parameter is generally used to assign the current process state. Preferably, several, and in particular a set of, process parameters are used. Each process parameter can take on values ​​within a predetermined range within the respective process state; that is, fluctuations of the process parameter within this range are permitted. By selectively influencing this at least one process parameter, it can, for example, be deliberately shifted into a middle range. This stabilizes the process parameter within its range. This also stabilizes the process state—provided such stabilization of the current process state is desired. Alternatively, if a state change occurs...

[0169] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 30 If desired, at least one process parameter is specifically controlled out of the value range and in particular controlled into a value range of a process state that is to be assumed.

[0170] Due to the high degree of interdependence between the various process parameters, this deliberate control is often non-linear and is preferably trained using advanced machine learning approaches based on historical data. The optimization measures are then dynamically adapted to the current plant process. Optionally, the operator can be shown the potential changes resulting from a control intervention or its omission, based on the stored models.

[0171] As previously explained, the plant process exhibits different process variants or can occur in different process variants. These are represented by the different model variants, as already described. The varying external influencing factors are those that cannot be determined or influenced by the plant or its control system during the plant process. Specifically, these are environmental conditions such as ambient temperature or humidity. In addition to environmental conditions, input variables or starting parameters, such as the raw material for the product being manufactured, also fall under these external influencing factors. For example, in the production of plastic products, the material used can be subject to fluctuations depending on the batch, which are reflected in different behavior of the plant process.Furthermore, changes to the system due to wear and tear are among such external influencing factors.

[0172] The method described here is therefore, in a preferred variant, carried out exclusively data-driven based on the determined process parameters during the ongoing plant process without external manual intervention. This means that no instructions issued by an expert (e.g., operating personnel), particularly process-related ones, are required.

[0173] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 31

[0174] Optimization instructions are implemented. These external interventions initiated by an expert are also referred to as expert-led interventions / measures.

[0175] In principle, there may be hard boundary conditions or empirical values ​​that are not adequately taken into account directly by this data-driven method.

[0176] In a preferred further development of the method, such expert-led measures are therefore provided, in particular according to claim 20. For this purpose, in particular a so-called expert system is implemented.

[0177] In particular, optimization measures are possible through manual intervention. Through these expert-led corrections / optimization measures, the model remains modifiable in the preferred training environment. Specifically, at least one specific assignment criterion for determining and assigning a respective process state (based on at least one process parameter) can be modified. For example, the expert can manually define and implement absolute hazard limits for certain process parameters in the model. If these limits are exceeded, this leads directly to a warning or even to the termination of the plant process. The assignment criterion for determining a respective process state refers, in particular, to the permissible value range, tolerance range, or distribution of a process parameter. This can therefore be manually set / modified by the expert system.

[0178] Another aspect of the expert system is the storage of optimization measures / recommendations for process parameters to achieve target values. These recommendations might include, for example, control or regulation instructions.

[0179] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 32

[0180] These expert-led optimization measures within the expert system therefore improve the overall quality and reliability of the monitoring and optimization of the plant process.

[0181] Expert-led measures, in this context, refer in particular to manual interventions by operating personnel.

[0182] Alternatively or additionally, expert-led measures also include automated measures, such as those initiated by an AI, based on external data sources—that is, data not generated by the plant process itself. Such external data sources include, for example, operating instructions, particularly in text form (manuals), data sheets, etc., which contain information about boundary conditions for process parameters and instructions for action. These external data sources can, for example, provide information on termination criteria and process end-times.

[0183] In preferred further training, the expert system can in turn be optimized in a data-driven manner by evaluating the expert-led measures according to their impact and feeding the results back into the expert system for optimization.

[0184] The expert system integrates expert knowledge, rules, heuristic relationships, and experiential knowledge into the modeling of process states and / or transition probabilities, particularly as a complement to data-based modeling. Alternatively, or in addition, these insights gained from the expert system are integrated into the model variants and / or classes.

[0185] By taking the expert system into account, especially for the further development of the model, a hybrid, adaptive combination of data-driven and expert-based improvement of the model is achieved.

[0186] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 33

[0187] The modeling, i.e., further development of the model, especially with the help of expert knowledge, preferably takes place continuously, so that the model, or the affected elements of the model, e.g. including the target variables, is / are continuously adapted and optimized.

[0188] Missing or incomplete data are preferably supplemented by simulations and / or expert rules. This provides, in particular, reliable forecasts, classifications, and / or recommendations for action.

[0189] From the relevance, direction and strength of process parameters for state changes, model or class assignments, concrete recommendations for action to achieve the goal can be derived.

[0190] The expert system is used in particular for the analysis and / or optimization of interconnected, temporally staggered, and / or parallel plant processes. For this purpose, global influencing factors and causal relationships are added using rule-based principles. This allows for the derivation of improved recommendations for action.

[0191] Linked plant processes, also known as parallel plant processes, are understood here to be similar or identical plant processes, specifically those processes used to manufacture the same or similar products. Linked plant processes are understood to be, in particular, processes that are executed in parallel at the same location, e.g., in the same hall, side by side, for example, on multiple production lines. Furthermore, linked plant processes also include those that utilize the same resources, such as the same batches of material.

[0192] Offline quality data, not collected during the plant process, is preferred for monitoring and optimization of the quality of the plant process or the manufactured product.

[0193] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 34 offline Quality data generally refers to data and information outside the actual plant process and thus to external data that contains information and statements about the quality of the manufactured product and thus implicitly also about the quality and excellence of the plant process.

[0194] This offline quality data refers in particular to data obtained from offline quality tests (laboratory tests) that accompany the plant process and / or to quality data from subsequent findings after completion of the plant process, such as product complaints.

[0195] Offline quality tests are typically used to regularly take and check samples of the manufactured products during the process, especially with regard to quality parameters that cannot be checked in the plant process.

[0196] The subsequent findings primarily concern complaints, returns, or defects in the manufactured products. In both cases, logging, for example using a production stamp (timestamp), makes it possible to assign the respective test result or subsequent product information to a specific time window within the plant process. The timestamp often offers only limited accuracy compared to the more frequently recorded process parameters. In this method, the plant processes are subdivided into longer time intervals or time windows based on the determined process states or process variants.

[0197] It is therefore possible to assign the insights gained to a specific process variant or, with sufficient temporal resolution, to a process state. Such an assignment is preferably made. Alternatively or additionally, an assignment to a respective model variant and / or a class of model variants is preferred.

[0198] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 35

[0199] The assignment is preferably made – as already explained – based on specific time intervals / time windows. Alternatively or additionally, the assignment is made based on the calculated probabilities of belonging to a specific process state, a specific model variant, or a specific class.

[0200] Offline quality data is preferably used for reassessment, supplementary evaluation, and / or optimization of process states, model variants, or classes. For example, a specific process state (model variant or class) might be classified as high-risk.

[0201] Classes, model variants, and / or process states associated with an increased probability of complaints and / or defects are preferably classified as higher risk. Therefore, risk indicators, which are primarily a measure of poor product quality, are preferably assigned to these classes, model variants, and / or process states based on offline quality data.

[0202] Based on these risk indicators, intervention in the plant process is preferably carried out and the plant process is appropriately adjusted. In particular, certain intervention measures and preferably also intervention times are changed / adjusted to avoid error-prone processes or products.

[0203] Preferably, testing measures such as test cycles are also changed depending on the offline quality data and especially depending on the risk indicators, so that, for example, more frequent testing / control takes place.

[0204] In the preferred configuration, if a current process course approaches a previously identified risky process course, a warning is automatically generated and / or a recommendation for action is issued.

[0205] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 36

[0206] By integrating and considering offline quality data, the particular advantage is achieved that previously unknown environmental or material influence classes can be systematically identified and their effects on the process can be quantified, so that targeted process adjustments and, in particular, warnings can be derived.

[0207] The insights gained from offline quality data are preferably incorporated into the model – similar to the insights obtained using or within the framework of the expert system. For example, these insights are used to assess the quality (quality characteristic) of a specific process state or process variant, or to evaluate and determine permissible value ranges for at least one process parameter. Thus, a process state previously considered acceptable or good can have its quality rating downgraded, potentially even to a critical state that may lead to a process stoppage to prevent scrap.

[0208] The plant process is preferably a continuous manufacturing process, in particular an extrusion process, in which the manufactured product is an extruded product.

[0209] Alternatively, the plant process is a discrete manufacturing process, in particular an injection molding process, and the manufactured product is an injection molded product, whereby the discontinuous process is treated as a quasi-continuous process by repeatedly performing several cycles.

[0210] The process parameters are selected from one or, in particular, several of the following parameters, such as

[0211] - the temperature, especially of the extrusion compound or the injection molding compound,

[0212] - the pressure, for example with which the extrusion compound is extruded, or the injection molding pressure,

[0213] - a rotational speed, for example of an extrusion screw,

[0214] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 37

[0215] - a residence time, for example of the injection-molded product in a mold,

[0216] - a geometric dimension of the manufactured product, e.g. the width, depth or dimensional accuracy of the manufactured product, etc.

[0217] - an optical feature, such as a surface structure, which is obtained, for example, through optical measurement / image capture, possibly with evaluation.

[0218] These process parameters are the main process parameters mentioned at the beginning. Secondary process parameters derived from these, such as mean, median, maximum, minimum, range, variance, slope, curvature, etc., of a respective main process parameter, are also considered process parameters.

[0219] For the creation, and especially for the preferably intended maintenance, continuous improvement, and training of the model, a sequence of state changes (between process states) is repeatedly learned and recorded in a preferred embodiment, so that a structure of the process state progression is obtained. The sequence is preferably determined over a complete manufacturing cycle, i.e., for example, from the start of production to the end of production, or until a termination occurs in the case of a termination criterion. At a minimum, the sequence extends over a sufficiently long process time, for example, at least one hour or at least several hours.

[0220] The transition probabilities between the different process states are determined based on the structure. The sequence between the various process states is used to determine the model's order and / or a probability for the transition from one process state to another. This probability calculation is then incorporated into the model's probability matrix. The determination of the model's order and / or the transition probabilities is carried out using well-known mathematical methods. For example...

[0221] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 38 so-called fit tests, such as the likelihood ratio test, are used.

[0222] Based on this structure, a residence time for a given process state, or a probability for such a residence time, is preferably also determined. This essentially establishes a measure for the residence time and survival time in a specific state. By considering the residence time, a time-weighted transition probability is implemented in the model; that is, the transition probability varies depending on the time / residence time that the plant process has already been in the same process state.

[0223] The structure of the process states refers in particular to the specific sequence of the various process states, especially the frequency of individual process states, their sequence, the frequency of state transitions, specifically which state transitions into which, and additionally, the duration in each state. The model can therefore be created based on these historical sequences or process structures.

[0224] Furthermore, the model is preferably continuously adapted and trained based on the ongoing plant processes.

[0225] The determination (modeling) of the model as well as the model variants and, in particular, the continuous training are preferably based on advanced machine learning or artificial intelligence methods.

[0226] As explained at the beginning, the method described here is based in particular on the fundamental assumption that there are hidden process states which are only manifested in the measured process parameters. For the creation and maintenance of the model, unknown model parameters of the (state) model are estimated based on the historically collected sequences of the defined process parameters. For this purpose, a suitable

[0227] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 39

[0228] Optimization, such as that used by the Baum-Weich algorithm, particularly employs a bidirectional approach. This bidirectional approach allows the model to consider the entire process flow with at least one forward and one backward pass. This results in the training of a comprehensive pattern structure, enabling the derivation of the emission probabilities of the underlying process states as well as their transition probabilities.

[0229] In preferred advanced training, the method described here is also used as a basis for simulation, particularly to model and simulate different processes and scenarios for the plant process. This includes simulating and modeling interventions and their effects on the plant process. Specifically, the process behavior is analyzed and evaluated in response to such interventions.

[0230] Model optimization:

[0231] In preferred further training, the model is continuously developed and, in particular, optimized. This involves the use of one or more different measures, which are explained in more detail below:

[0232] Model optimization through transformation / meta-states:

[0233] Process states are preferably used not only in their originally defined form, but also transformed, summarized, or abstracted. Expert knowledge is particularly utilized for this purpose.

[0234] Preferably, meta-states such as process capability, process failure or (process) end states such as end of production or process termination due to machine or process errors are defined and taken into account in the model.

[0235] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 40

[0236] This enables reduced model complexity in favor of purely process capability-oriented optimization and results in better overall process control.

[0237] Model optimization through covariables and KPis:

[0238] The model is preferably optimized using additional, explanatory parameters (so-called co-variables) and / or further information. These co-variables are, in particular, the secondary process parameters described above, i.e., quantities derived from the process parameters such as mean, median, minimum, maximum, slope, curvature, range of variation, and variance. An example of such a co-variable is, in particular, average energy consumption. For instance, one process state might show high energy consumption with high product quality, while another process state might show lower energy consumption with lower product quality.

[0239] The co-variables therefore include additional relevant process parameters and / or derived key performance indicators (KPIs) that correlate with the process states and, for example, explain their variance. These co-variables are considered particularly for optimizing the process states under defined constraints.

[0240] For modeling the process states, a normal distribution is preferably assumed. Alternatively or additionally, non-normally distributed, multimodal, discrete, or empirically determined distribution functions are also used. Their selection is primarily data-driven and / or based on expert specifications.

[0241] In addition to data-based transition probabilities, rule-based and / or knowledge-based relationships are also given preference.

[0242] Overall, improved, more realistic and more robust process optimization is made possible, in particular through a more precise description of the process states with regard to competing target variables.

[0243] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 41

[0244] Material used:

[0245] The plant process is, in particular, a manufacturing process for producing a product, in which a starting material with a predetermined material composition, especially plastic or metal, is used. Within the framework of sustainability aspects, a proportion of recycled material is often added to the material composition, e.g., to the material (granules) fed into an extruder or injection molding machine; that is, a recycled content is preferably used for the starting material.

[0246] The method described here is used in particular to make a statement about the quality of the material composition used, i.e. to draw conclusions about the quality of the material composition used.

[0247] Preferably, the material composition and in particular the recycled content in the model is taken into account, and preferably also during the modeling process, i.e., during the creation, optimization and further development of the model.

[0248] Preferably, changes in the process flow are analyzed, in particular based on a changed assignment to process states, model variants and / or classes, especially with regard to an influence of the material composition.

[0249] A change in classification (state change), for example to a process state or to a class with lower quality characteristics, is preferably used as a warning signal. Such a change in classification should therefore be considered an indication that, for example, a less suitable material composition is being used.

[0250] Therefore, a statement about the quality of the material composition is made based on such a change in assignment / change of state.

[0251] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 42

[0252] Preferably, a change in the process, which for example exceeds a predetermined threshold, is used as an indicator of the quality of the material composition (material batch quality), in particular of the recycled material used.

[0253] In preferred further development, the material composition, especially the recycled content, is dynamically optimized, particularly in such a way that the most sustainable material use possible is achieved while adhering to predefined tolerances and target parameters (quality, throughput, energy efficiency). Thus, the material composition is preferably modified while the process is continuously monitored, especially with regard to changes in state.

[0254] The preferred method for evaluating the material quality of the starting material, particularly the recycled material, is described here. For this purpose, material compositions (material batches) that do not exhibit adverse effects, such as process drift, even at higher proportions, and that result in consistent or improved quality characteristics, are classified as correspondingly high-quality. Material compositions (material batches) that lead to disadvantages, such as process drift or deteriorated quality characteristics, even at low proportions, are classified as inferior.

[0255] Cross-process optimization.

[0256] In preferred advanced training, the procedure described here is not only applied to a single plant process, but also to several plant processes that are correlated and specifically linked together.

[0257] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 43

[0258] Interrelated plant processes are understood to be those processes that are linked, as previously defined. Interrelated plant processes also include processes that are staggered in time, run in parallel, and / or span multiple locations. For example, interrelated plant processes also encompass different process types, particularly for the production of different products. Preferably, however, the interrelated plant processes exhibit the same process types / process sequences and / or serve, in particular, to produce the same products.

[0259] Preferably, the previously described assignment to a class is carried out across processes. This enables a coherent optimization of multiple plant processes.

[0260] Through cross-process analysis, i.e., an analysis of the correlated plant processes, and in particular an analysis of the assigned classes, model variants and / or process states, cross-process influencing factors are identified in preferred further training, which result in particular from the shared use of process facilities, material compositions or site-specific environmental influences (e.g. cooling systems, air currents, heat load) and are recognized based on similar process sequences.

[0261] This analysis preferably distinguishes between process-specific and cross-process influencing factors, so that appropriate measures can be derived and non-process-specific inefficiencies can be optimized or avoided across processes.

[0262] In measures derived across processes, the specific characteristics of each individual plant process are given priority. b) Process transfer:

[0263] Preferably, based on the cross-process analysis and in particular through the assigned classes, model variants and / or process states, a transfer of process parameters, model components, intervention strategies and

[0264] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 44

[0265] Root cause analyses of one plant process affecting another plant process are preferably carried out based on identified similarities.

[0266] This transfer occurs particularly when certain influencing factors or process parameters are unknown, not collected, or not measurable in the other process.

[0267] Systematic parameter integration or imputation from other plant processes preferably continues to support robust modeling and process optimization, even for new or incomplete processes. c) Prioritization

[0268] Cross-process analysis preferably prioritizes intervention measures and / or intervention times for multiple parallel, interconnected, or site-spanning processes. This enables orderly and resource-efficient optimization, particularly when deviations occur simultaneously in several machines or plant processes, and / or prevents interruptions, downtime, or operator errors due to overload.

[0269] With regard to the technical system according to claim 27, a preferred embodiment comprises a system complex consisting of several individual technical systems (individual systems). These individual systems are, for example, distributed across different locations (different buildings, different towns) or housed in a common location, such as a shared plant hall. For instance, several production lines or assembly lines each constitute an individual system. In particular, the multiple individual systems are monitored jointly. This allows, for example, the derivation of common recommendations for action for all individual systems if several or all individual systems exhibit a specific behavior. For example, if the individual systems show a temperature drift, a common recommendation for action, such as lowering the temperature in the plant hall, can be issued for all individual systems.These comprehensive measures, for example, prevent isolated symptom treatment limited to individual systems.

[0270] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 45

[0271] Deviations were avoided, the cause of which likely lies in other, higher-level process influences.

[0272] Exemplary embodiments of the invention are explained in more detail below with reference to the figures. These show simplified representations of:

[0273] FIG 1 shows a simplified representation of an extrusion plant as a technical system for carrying out an extrusion process as a plant process.

[0274] FIG 2 a probability matrix,

[0275] FIG 3 shows a table with process parameters and their assignment to a process state.

[0276] FIG 4 shows a representation of the assignment of different process variants to corresponding model variants,

[0277] FIG 5 Illustrations to explain the adjustment dimensions for different model variants, as well as

[0278] FIG 6 shows a diagram to illustrate the selection of the prioritized model.

[0279] Figure 1 shows a highly simplified extrusion plant as a technical system 2 for carrying out an extrusion process as a plant process for the production of a product 4. In this exemplary embodiment, the product 4 is a continuously produced (endless) product, for example, an extruded profile or a so-called profile strip, which is used, for example, as a decorative edge banding for furniture. A crucial quality characteristic of such profile strips is, in particular, the width B of the profile strip.

[0280] Technical system 2 generally comprises a production machine, in this case an extruder 6, to which starting material A, for example in the form of plastic granules, is fed. In the case of an injection molding system, the production machine is an injection molding machine.

[0281] The starting material A is heated and melted by the extruder 6 and finally, using a screw, through an extrusion die to produce

[0282] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 46 of the continuous product 4 is pressed out. This is wound up, for example, after a cooling section.

[0283] Technical system 2 further comprises a control device 8, which controls and regulates the system 2. Various process parameters are continuously recorded during the system process via several sensors S1 to S4. These process parameters P include, in particular, the (extrusion) pressure p, the (extrusion) temperature T of the material, and preferably the rotational speed n of the screw of the extruder 6. Furthermore, in the exemplary embodiment, a sensor S4 is provided for detecting at least one geometric dimension or other optical characteristic of the extruded product 4, in particular the width B. The instantaneous energy consumption E is also determined, for example, based on an instantaneous current draw.

[0284] The control device 8 is used, in particular, to adjust the temperature T, the pressure p, and the rotational speed n using suitable, known methods. The values ​​for the various process parameters P, as recorded by sensors S1 to S4, are transmitted to the control device 8. This device can also be divided into several physical units and, for example, may also include one or more connected external computing units. The plant process is monitored based on these process parameters P, and preferably, the plant process is also controlled, regulated, and optimized, particularly automatically.

[0285] For monitoring and, in particular, optimization, a model M of the plant process, especially in the form of a digital twin, is stored within the control device. Preferably, different model variants M1 ... Mn are stored for different process variants of the plant process.

[0286] Based on the current temporal progression of the process parameters P (more precisely, the values ​​of the process parameters P), a current progression V of the plant process is recorded.

[0287] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 47

[0288] As will be explained in detail below, the current process state V, i.e., the current process parameters P, is continuously assigned to a current process state Z. A process state is generally denoted by the reference symbol Z, with the individual process states being distinguished by the reference symbols Zi, Zj, or specifically Z1 ... Zn.

[0289] The model Mx (selected model variant) includes, among other things, a probability matrix W, as illustrated in Figure 2. This probability matrix W contains transition probabilities Wij, where each transition probability Wij represents the probability of a state change from process state Zi to process state Zj. Preferably, this is not a static transition probability Wij but a time-dependent transition probability Wij(t), which takes into account the residence time in each process state. This residence time was previously considered in the model Mx and, for example, stored as a time-dependent function (hazard rate) for each state change. This allows the instantaneous risk of the plant process leaving the current process state Z within a specific time interval to be determined with time precision.

[0290] The probability matrix W is in particular a higher-order probability matrix W in which the transition probability Wij for a respective change of state from a process state Zi to a process state Zj depends not only on the current process state Zi, but also on at least one or more of the preceding process states Z.

[0291] In the probability matrix W shown in FIG. 2, process states Z2, Z3, Z4, and Z5 are given as examples of different predecessor process states for a process state Z1 (leftmost column). The various probability matrix W can be used to determine the different

[0292] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 48

[0293] Transition probabilities Wij (t) to the other process states Zi starting from process state Z1 are given. The sum of each row adds up to 1.

[0294] Process states Z1 to Z5, for example, are represented as characterizations of normal process states Z1 within permissible tolerance fields for the process parameters P. Process state Z6, for example, is a final state and a termination state resulting from a machine failure. Process state Z7, for example, is a final state and a termination state resulting from, for example, a product or process failure, if one of the process parameters, such as temperature T or width B, or a secondary process parameter derived from this main process parameter T or B, such as the variance of temperature T or width B, exceeds a permissible range of values, in particular a tolerance range. The last process state, Z8, defines in particular a final state at the end of production, i.e., if no disturbance has occurred and, for example, a predetermined production quantity has been reached or a change of order has taken place, etc.

[0295] FIG. 3 shows an example and partial representation of a table with continuously recorded process parameters P. For each time interval, which in this case covers a period of 10 seconds, a process state is assigned based on the process parameters P. The columns in the table therefore indicate a timestamp Ts, the width B, the pressure p, the temperature T, the screw speed n, and the energy consumption E. Each of these process parameter values, recorded at a defined time, is assigned a process state Zi. Each row thus defines a data record for a respective time interval. Such a data record can also contain metadata, for example, about the extruder used, the article number of the manufactured product, and other additional information.

[0296] For example, an aggregated data set is generated as a further data set, while maintaining the order of the individual process states, in which

[0297] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 49 For example, based on the measurement data for a respective main process parameter P, derived parameters such as minimum, maximum, mean, slope, variance, etc., can also be used. In particular, an actual residence time t is also determined, i.e., the duration for which the plant process remains in a respective process state.

[0298] The assignment to a respective process state Z is based on at least one of the process parameters P and preferably on a whole set of process parameters P.

[0299] Preferably, no fixed, static assignment to a respective process state Z is made based on fixed range values ​​for a given process parameter. Instead, an assignment probability is used.

[0300] For example, for a process parameter P, the mean of a range of values ​​is associated with a high probability, while the extreme values ​​are associated with a low probability, indicating a specific process state Zi. Therefore, over a given range of values ​​for the process parameter P, the probability that a specific process state Zi is assigned follows, for example, the pattern of a given mathematical function, in particular a normal distribution around the mean of the range. This probability distribution is also referred to as the distribution probability.

[0301] The probability of assignment is ultimately obtained through a combination, in particular through the product of this distribution probability with a (fixed) transition probability.

[0302] It is preferred that the relevance, and preferably also the (local) direction and strength of influence of the respective process parameters P, be stored for the assignment to a specific process state Zi. For example, the mean pressure value determined within the process state or

[0303] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 50 The temperature has a high relevance for the assignment to a respective process state Zi, whereas, for example, a determined minimum or maximum of the temperature or pressure within the process state has only a low relevance and thus little influence.

[0304] For the optimization of the plant process, particular attention is paid to this information regarding the relevance, and preferably also to the (local) direction and strength of influence of the process parameters P, as well as additionally preferably to the probability of assignment, which is also time-dependent.

[0305] In this way, either to stabilize the current process state Z or to deliberately switch to another process state Z, one or more of the process parameters are specifically influenced in order to stabilize the current process state Zi or to switch to another process state Zj.

[0306] Due to varying external influencing factors, such as varying environmental conditions (temperature fluctuations, humidity fluctuations, etc.) or also due to varying input variables - such as a varying starting material A - the (current) course V of the plant process varies, so that depending on the varying external influencing factors, which are often not identifiable or measurable, the plant process takes place in different process variants V1 .. Vn.

[0307] These different process variants V1 ... Vn are assigned to the different model variants M1 ... Mn in model M, as illustrated in FIG. 4. Based on the current course V1 ... Vn of the plant process V, the most suitable model variant M1 ... Mn is assigned by comparison with the different model variants M1 ... Mn and used for further monitoring, control, and regulation of the plant process.

[0308] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 51

[0309] For practical purposes, the different model variants M1...Mn and, correspondingly, the different process variants V1...Vn are classified and assigned to different classes K1...Km. For example, several process variants V1...Vn or several model variants M1...Mn are assigned to a common class K1...Km. The individual classes K1...Km can also be further subdivided into subclasses.

[0310] The assignment to a respective class K1 ... Km (or subclass) is based on characteristic values ​​for the behavior of at least one process parameter P and / or the process states Zi. Characteristics include, for example, the frequency and number of jumps in the values ​​of the process parameter P or the number of state changes Zij. Alternatively, the behavior can exhibit a characteristically curved pattern, for example, in the form of a sinusoidal oscillation. Another characteristic feature is, for example, a continuous increase in the value of the process parameter P, for instance, as a result of a so-called drift. Such a characteristic indicates that the current plant process and its behavior are critical and that countermeasures may be necessary. Furthermore, a characteristic feature can consist of a largely constant value for the process parameter, which, for example,within small stochastic fluctuations. This indicates, for example, a stable process state.

[0311] Each of these classes K1.. Km is expediently assigned a quality indicator, for example in the form of a color code (green - yellow - red), which is also displayed to the user, specifically within the framework of an assistance system. The assistance system generally serves to support the operator / plant manager by visualizing the plant process and / or providing recommendations for control interventions.

[0312] During monitoring and preferably automatic control to optimize the plant process, the process parameters P are continuously recorded and continuously adjusted based on at least one of these recorded parameters.

[0313] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 52

[0314] The process parameter P is assigned to a current process state Zi.

[0315] Based on the stored probability matrix W, the transition probability Wij for a state change to another process state Zj is determined, starting from the current process state Zi. Different quality indicators are assigned to the different process states, derived, for example, from the values ​​of the process parameter P, which are relevant for the assignment to the respective process state Zi.

[0316] For example, at least one of these process parameters P is determined, particularly manually, as a target variable X to be achieved. For instance, this target variable X might be the width B of the manufactured product 4. A small range of variation may be desired for high quality, or a wider range for lower quality. With lower quality requirements, for example, minimized energy consumption E within the permissible tolerance range might be the primary target variable, or indeed the target variable X. That is, with the target variable X "low energy consumption E," a larger range of variation is permitted for the parameter value of width B.

[0317] Depending on the selected target variable X, the system uses transition probabilities to determine recommended actions and displays them, for example, visually or initiates them automatically. If the target variable X is lower energy consumption, which is achieved, for example, in process state Z3, the recommended action is to achieve or maintain process state Z3 as much as possible. Based on this, and considering the knowledge gained from model Mx, particularly the relevance and, in particular, the (local) direction and strength of influence of the various process parameters P, the system then initiates a specific action, for example, to change one or more operating parameters P.

[0318] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 53

[0319] A particular advantage of the method described here is that, by assigning processes to current process states Zi, which are essentially classified, and by considering the transition probabilities Wij to other process states Zj, taking into account experience and histones, potential errors or poor quality can be detected early within the framework of continuous process monitoring, and countermeasures can be taken in a timely manner.

[0320] Advantages of the method described here also arise in particular from the possibility of optimization control, for example to consciously reduce resource consumption (energy, material input).

[0321] Furthermore, it is worth highlighting the possibility of preventing or at least sorting out potentially poor-quality products early on, or at least in time, through offline data processing (e.g., evaluation of plant processes and / or the progression of selected process parameters), thus preventing later complaints, for example.

[0322] In particular, the assistance system preferably continues to present the plant process in a compressed and understandable / intuitive way using a so-called digital twin based on the stored state model.

[0323] Furthermore, the method described here, with its underlying state model M – particularly in conjunction with the various model variants M1 ... Mn – offers the possibility of realistically simulating the plant process. This includes simulating different process progressions / scenarios and, in particular, simulating or investigating the effects of various control interventions. Such simulations are preferably performed. The insights gained from these simulations are then used, for example, for controlling the actual plant process and for further developing the underlying model M.

[0324] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 54

[0325] Figures 5 and 6 describe measures for assigning a model variant M1 — Mn to the current course of the plant process, i.e., to the current course V1.. Vn (process variants) of the plant process:

[0326] Preferably, in a first step, the behavior of one or more process parameters P (for example, a characteristic value (quality parameter) of a manufactured product, such as a desired geometric dimension) is determined using a large number (preferably more than 100, and especially more than 500) of simulations with the different model variants M1..Mn – under given boundary conditions. Thus, a large number of individual behaviors are simulated for each model variant M1 ..Mn.

[0327] Therefore, in these simulations, a process progression results for each simulation and each model variant M1.. Mn.

[0328] In FIG 5, the left half of the image shows the different curves of the process parameter P of the individual simulations over a given time range AT in a superimposed representation for different model variants M1-M3.

[0329] From this multitude of scenarios, a multitude of fit measures a are derived. For this purpose, each simulated scenario is compared with every other simulated scenario, and a fit measure a, in particular a difference measure, is determined for each, as described in the introductory section.

[0330] From these determined adjustment measures a, a frequency distribution of the values ​​of the adjustment measure a for the respective model variant M1.. Mn is preferably determined. The resulting frequency distributions are shown for each model variant on the right half of FIG. 5.

[0331] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 55

[0332] In the exemplary embodiment of model M1, the adjustment factor a varies between -75 and -145, in model M2 between 10 and 70, and in model M3 between -800 and -400. In the illustrated embodiment, the adjustment factor a was determined using the likelihood function, in particular the log likelihood function, which is known per se.

[0333] It should be emphasized that the fit measure a, and in particular the resulting frequency distribution, is time-dependent. Specifically, the fit measure a, and thus the resulting frequency distribution, changes depending on, for example, a considered time window / period (e.g., 100s, 500s, 1000s) for the respective process.

[0334] Preferably, for different time windows (e.g., for more than 3, more than 5, or more than 10 different time windows), a frequency distribution (dependent on the chosen time window) is determined for each of the different model variants (by means of a large number of simulations).

[0335] During an ongoing plant process, its instantaneous profile V (for a given time window) is recorded. Figure 6 illustrates an example of an instantaneous profile V (for a time window of 100) in the upper half of the image.

[0336] This current trend is compared with a trend simulated using the respective model variant.

[0337] During the ongoing process, the adjustment measures a are continuously determined for each model variant M1 .. Mn. For this purpose, the current, actual curve V is compared with a currently simulated curve, and the adjustment measure, e.g., the difference measure or using the likelihood function, is determined from this comparison. This constitutes the current, actual adjustment measure a (for each model variant M1.. Mn).

[0338] Based on the previously performed simulations, a normalized value is calculated for the current adjustment measure a of the respective model variant M1 ... Mn.

[0339] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 56

[0340] The probability H(a) is assigned, which is also called the assignment probability. This is determined in particular from the frequency distribution described above. Thus, it considers the probability with which the instantaneous fit measure a occurred in the simulated scenarios.

[0341] The table shown in FIG 6 below lists the instantaneous fit measures a of the different model variants M1.. Mn and their normalized probabilities H(a) (assignment probability).

[0342] Based on this evaluation, the model variant M1 ... Mn is then assigned and selected (i.e., assigned to the current course V and thus selected as the model variant (prioritized model) used to determine the process states and their transition probabilities) where the fit measure a has the highest probability H(a). In the case of FIG. 6, this is model variant M1.

[0343] It should be emphasized that the model variant M1 .. Mn is selected where the determined instantaneous, actual distance measurement has the highest probability of assignment. This is not necessarily the model variant with the best fit.

[0344] In the embodiment of the table according to FIG. 6, the highest assignment probability H(a) happens to coincide with the best fit measure a for model variant M1. However, a comparison of model variants M2 and M3 already shows that, for example, the fit measure a is better for model variant M2 than for M3, but the assignment probability H(a) is significantly worse. Therefore, if a choice had to be made between the two model variants M2 and M3, model variant M3 would be selected as the prioritized model despite its worse current, actual fit value a.

[0345] Therefore, the preferred decision criterion for selecting which model variant M1...Mn is chosen is preferably not the value of the fit measure a, but rather its probability H(a).

[0346] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 57 results from the assignment to the respective distribution determined from the simulations.

[0347] In a preferred embodiment, this verification and adjustment of the adjustment parameters a is carried out continuously or at predetermined time intervals, but in all cases in real time and / or process-accompanying during an ongoing plant process.

[0348] The same time windows are used for both the current actual course and the simulated courses and the frequency distributions derived from them.

[0349] This takes into account temporal changes and variations in the adjustment measure a. These arise in particular from potentially altered influencing conditions and process dynamics.

[0350] Furthermore, a change between the model variants M1.. Mn preferably takes place automatically during the ongoing plant process based on these, in particular, time-varying adjustment measures a, as described in FIG 6.

[0351] Therefore, if it turns out that the currently selected model variant describes the current course V worse than another model variant, this other model variant will be selected for the further control of the plant process and in particular for determining the process states and their transition probabilities.

[0352] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 58

[0353] Reference symbol list

[0354] 2 technical system

[0355] 4 products

[0356] 6 extruders

[0357] 8 Control device

[0358] A starting material

[0359] P Process parameters

[0360] S1-S4 sensors

[0361] B width

[0362] T Temperature p Pressure n Speed

[0363] Energy consumption

[0364] M Model

[0365] M1..Mn model variants

[0366] V current course of the plant process

[0367] V1.. Vn process variants

[0368] Z, Z1.. Zn Process state

[0369] W probability matrix

[0370] Wij transition probability

[0371] Ts timestamp t dwell time

[0372] X target variable a adjustment measure

[0373] H(a) normalized probability for adjustment measure a

[0374] (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025

Claims

1. FDST Patent Attorneys, Nuremberg Page 59 Claims 1. Method for monitoring and controlling a technical plant process, in particular an industrial manufacturing process for producing a product, in which - a time course of at least one process parameter (P) such as pressure, temperature, geometric and / or optical characteristics as well as energy consumption is recorded, - based on at least one process parameter (P), a current process state (Zi) is continuously determined, so that a progression of process states (Zi) is obtained, wherein each process state (Zi) is a key figure for the quality of the plant process or of a product manufactured with the plant process, wherein preferably a final state of the plant process is used as a further process state (Z7, Z8), - a trained model (M) for the plant process is used, which contains a probability matrix (W) that specifies probabilities (Wij) for a change of state (Zij) between different process states (Z), - using the model (M) a transition probability (Wij) for the change of state from the current process state (Zi) to another process state (Zj) is determined, - based on the current process state (Zi) and the determined transition probability (Wij), a measure is initiated to achieve a predetermined target variable (X) during the execution of the plant process.

2. Method according to the preceding claim, wherein the plant process is subject to fluctuations as a result of varying external influencing factors, namely in particular environmental conditions and / or input variables such as, for example, the starting material (A) for the product to be manufactured, such that (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg Page 60 the plant process can occur in different process variants (V1 ....Vn)), whereby different model variants (M1 ... .Mn) are stored for the different process variants (V1 ... .Vn).

3. Method according to the preceding claim, wherein the different model variants (M1 ....Mn) are compared with the current plant process and the model variant (M1 ....Mn) that best represents the current plant process is selected as a prioritized model (M) and is used as the trained model (M) of the plant process.

4. Method according to the preceding claim, wherein during the current plant process a switching between the model variants (M1.. Mn) takes place, in particular automatically, depending on at least one adjustment measure (a).

5. Method according to one of the two preceding claims, wherein, for the selection of a model variant (M1 ... Mn) as the prioritized model (M), a distribution of at least one adjustment measure (a) for a respective model variant (M1 ... Mn) is taken into account, wherein the distribution is determined in particular by a multitude of simulations of the plant process with a respective model variant (M1 ... Mn).

6. Method according to the preceding claim, wherein a probability for the respective adjustment measure (a) is derived on the basis of the distributions and the probability for the selection of a model variant (M1 ... Mn) is used as the prioritized model (M).

7. Method according to any one of claims 4 to 6, wherein the adjustment measure (a) is time-dependent and is in particular a time-progressive adjustment measure (a).

8. Method according to one of claims 2 to 7, wherein the different model variants (M1 ....Mn) are classified into different classes (K1 ... Km) based on characteristic features of the course of the plant process. (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg, page 61, are classified, with each class (K1 ... Km) preferably being assigned a quality indicator.

9. Method according to one of the preceding claims, wherein a final state of the plant process and in particular a termination criterion of the plant process is used as a further process state (Z7, Z8).

10. Method according to one of the preceding claims, wherein the measure to achieve the predetermined target size (X) is initiated as soon as a value for the at least one process parameter (P) remains within a permissible tolerance range.

11. Method according to one of the preceding claims, wherein information of the plant process is displayed via a user interface and, in particular, interventions in the plant process are also made possible via the user interface.

12. Method according to one of the preceding claims, wherein the target variable (X) is optionally at least one predetermined quality parameter for the plant process or for the product, a material input for the manufacture of the product, and / or an energy consumption of the plant process or for the manufacture of the product, wherein the target variable (X) is preferably manually selectable.

13. Method according to one of the preceding claims, in which several target variables (X) are taken into account and a prioritization and in particular selection of a target variable (X) as the predetermined target variable (X) is carried out, wherein the prioritization is preferably carried out dynamically.

14. Method according to one of the preceding claims, wherein for the determination of the transition probability (Wij) a previous course of the transition between previous process states (Z) is taken into account. (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg, page 62 15. Method according to one of the preceding claims, wherein the residence time (t) in the current process state (Zi) is taken into account for the determination of the transition probability (Wij), wherein the transition probability is preferably represented as a function of the residence time.

16. Method according to one of the preceding claims, wherein the assignment of a momentary process state (Zi) on the basis of the at least one process parameter (P) is carried out using an assignment probability.

17. Method according to one of the preceding claims, wherein, for the determination of a change of state (Zij) between two process states (Zi, Zj) based on a change of the at least one process parameter (P), the transition probability (Wij) for the change of state between the two process states (Zi, Zj) is taken into account.

18. Method according to one of the preceding claims, wherein the relevance, and preferably additionally a direction of influence and / or an intensity of influence of different process parameters (P) for the change of state (Zij) between two process states (Zi, Zj) is stored in the model.

19. Method according to one of the preceding claims, wherein the at least one process parameter (P), according to which the current process state (Zi) is assigned, can assume a predetermined range of values ​​for the respective process state (Zi), wherein, within the framework of an optimization measure for the plant process, the at least one process parameter (P) is specifically changed in order to stabilize the current process state (Zi) or to specifically bring about a different process state (Zj).

20. A method according to any of the preceding claims, wherein expert-led measures are permitted within an expert system, wherein in (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg, page 63, preferred training programs - through expert-led measures, the model (M) is continuously optimized and / or - the expert-led measure includes recommendations for action to achieve the objective and / or - the expert-led measures are continuously recorded and evaluated, particularly with regard to their impact, and the expert system is modified based on these evaluations.

21. Method according to one of the preceding claims, wherein offline quality data for the quality of the plant process or the manufactured product, which are not recorded during the plant process, are used for monitoring and control, wherein the offline quality data optionally or in combination are offline quality tests accompanying the plant process or subsequent findings after completion of the plant process, such as product complaints.

22. Method according to the preceding claim, wherein the offline quality data is assigned to a process state and / or a model variant according to any one of claims 2 to 7 and / or a class according to claim 8.

23. Method according to the preceding claim, wherein a risk indicator is assigned to the respective process state and / or the respective model variant or the respective class based on the offline quality data, wherein intervention in the plant process is carried out depending on the risk indicator.

24. Method according to one of the preceding claims, wherein a sequence of state transitions (Zij) between process states (Zi, Zj) is repeatedly recorded, such that a structure of the course of the process states (Z) is obtained, wherein the transition probabilities (Wij) for storage in the model (M) are determined on the basis of the structure. (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025 FDST Patent Attorneys, Nuremberg, page 64 25. Method according to any of the preceding claims, which is used for simulation to analyze and evaluate possible scenarios, interventions or process behavior.

26. Method according to one of the preceding claims, wherein the plant process is a manufacturing process for producing a product in which a starting material (A) with a predetermined material composition is used, wherein in particular a recycled content is used for the starting material, whereby the quality of the material composition is inferred.

27. Technical installation (2) which is set up to carry out a method according to one of the preceding claims. (\\fs2012\gsi-software\winpat5\document\amt\4121849 docx) Last saved: December 10, 2025

Images