Method for categorising and method for monitoring a production in a paper machine, in particular web breaks
By categorizing production disruptions on paper machines using historical signal data and machine learning, the method provides reliable predictive models to prevent web breaks and quality deviations, improving productivity and reducing downtime.
Patent Information
- Authority / Receiving Office
- EP · EP
- Patent Type
- Patents
- Current Assignee / Owner
- VOITH PATENT GMBH
- Filing Date
- 2021-09-01
- Publication Date
- 2026-06-17
AI Technical Summary
Existing methods for predicting web breaks on paper machines are complex, require significant data preparation, and provide unreliable predictions, leading to unpredictable production disruptions and high costs.
A method that categorizes production disruptions by analyzing historical signal data from a paper machine's process and quality control systems, grouping similar disruptions into categories, and creating category-specific predictive models using machine learning to identify and prevent future disruptions.
The method effectively identifies and groups similar production disruptions, allowing for timely and targeted adjustments to prevent web breaks and other quality deviations, enhancing productivity and reducing downtime.
Smart Images
Figure IMGF0001 
Figure IMGF0002 
Figure IMGF0003
Abstract
Description
[0001] The invention relates to a method for categorizing production disruptions on a paper machine, in particular web breaks, and to a method for monitoring the production of a paper machine, in particular web breaks, using a method for the aforementioned categorization. It further relates to a computer program for executing one of these methods on a computer unit.
[0002] Here, the term "paper machine" refers to all machines used for the production and / or processing of fibrous webs, including not only machines for the production of paper webs but also, for example, so-called cardboard machines, pulp machines, coating machines or paper processing machines.
[0003] On paper machines, fibrous webs, some with very low tensile strength, are transported through the machine. Despite highly developed plant technology, production disruptions occur repeatedly, interrupting smooth production, particularly web breaks. This results in undesirable downtime and high costs. Besides web breaks, a production disruption can also be a quality deviation that falls outside the permissible limits for product quality, thus interrupting smooth production. In many cases, the paper machine must also undergo extensive cleaning after a production disruption before production can resume. Such production downtimes incur high costs and must therefore be kept to a minimum. These disruptions, and especially web breaks, are usually unpredictable events for operators.
[0004] While there are general measures, such as regular preventative cleaning of the system and more stable regulations to reduce the frequency of malfunctions, there are currently no reliable methods to proactively avoid malfunctions, especially track breaks.
[0005] Prior art methods exist that use data analysis to indicate an increased risk of web breaks. However, these methods are very complex and still unreliable, and not yet sufficiently developed to truly prevent web breaks. For example, US Patents 2002 / 0052699 A1 and 2002 / 0038197 A1 disclose methods for predicting web breaks on a paper machine. Using sophisticated algorithms, based in part on regression analyses of datasets from previous breaks, the method determines a break sensitivity and predicts the time until the next break. The goal is to create a predictive model that can warn of future breaks.One disadvantage of these well-known methods is that they require a high level of effort to filter and clean the data before regression analyses can even be applied, and that they still only produce inaccurate predictions.
[0006] The object of the invention is to develop a method that contributes to, or is suitable for, preventing production disruptions, in particular web breaks on a paper machine, with a higher probability, and thus leading to higher productivity of the paper machine. Furthermore, the object is to provide a corresponding computer program for carrying out this method.
[0007] The problem is solved by a method according to claim 1 or by a method according to claim 8. Furthermore, it is solved by a computer program product according to claim 7 or claim 13. Further advantageous features are specified in the respective dependent claims.
[0008] According to the invention, the method according to claim 1 is characterized in that the following steps are carried out: 1.1 Providing an initial data set (D1) of historical signal values from numerous different signals (S1...Sx) from a paper machine production system as time series, wherein the signals originate at least partly from a process control system (DCS), in particular from a drive control system, and at least partly from a quality control system (QCS) of the production system. 1.2.1 Selecting a fault point in time at which a production fault, in particular a web break, occurred; for a plurality of the signals from the initial data set, in particular for each signal, determining the change (ΔSx...ΔSz) of the respective signal value in a time interval before the selected fault point in time; and assigning a characteristic numerical value (ZW1...ZWx) to each signal change, wherein this numerical value (ZW1...ZWx) is higher the more significant the signal change, and wherein each numerical value (ZW1...ZWx) is assigned to a signal (S1...Sx).1.2.2 Summarize the individual characteristic numerical values (ZW1...ZWx) for the selected disruption time into a series (R1...Rn). 1.3 Repeat steps 1.2.1 and 1.2.2 several times, each time for further selected disruption times (t1...tn), so that further series (R1...Rn) of characteristic numerical values (ZW1...ZWx), assigned to the respective disruption times, are formed. 1.4 Segment the series (R1...Rn) thus obtained into different categories (K1...Km) by assigning similar series (R1...Rn) to the same categories (K1...Km), whereby series (R1...Rn) are considered similar if they show substantial similarity in the characteristic numerical values (ZW1...ZWx); each category (K1...Km) represents a type of production disruption, in particular a type of web break. In particular, series that cannot be assigned to any category are added to a residual set. 1.5 Output of the generated categories (K1... Km) and / or the series assigned to the respective category (R1... Rn) to an end device (24) and / or in the DCS and / or QCS. 1.6 Formation of a predictive model (P1... Pm) for each of the categories (K1... Km) by evaluating the series (R1... Rn) assigned to the respective category (K1... Km), whereby a separate predictive model (P1... Pm) is formed for each category, such that the predictive models (P1... Pm) are suitable for evaluating a second data set of signal values as time series and specifying a significator for the associated category (K1... Km) that describes a category-specific interference risk.
[0009] The steps do not necessarily have to be performed in the specified order, but can also be carried out in a different order. Furthermore, individual steps can also be performed in parallel.
[0010] In particular, the repetitions of steps 1.2.1 and 1.2.2 can be performed in parallel with the first execution and do not have to be performed sequentially.
[0011] A key advantage of the method according to the invention is that it allows similar production disruptions to be identified and grouped into a category. Additionally, a separate predictive model is created for each category of production disruption, which can provide a significator for the category-specific disruption risk. This provides the production control system (DCS) and / or the operator with important information for improving the adjustment of production parameters, as a specific risk of occurrence can be determined for each category of production disruption, particularly for each type of web break. This helps to prevent future production disruptions through timely adjustments to production settings. In particular, the method offers the possibility of using production system data without complex synchronization or data cleansing.Signal values that are not needed or usable for categorization are simply not used further. In particular, they can be assigned to a residual set.
[0012] The signals provided in the first data set from the production system can originate from sensors or measuring devices, for example. They can be control signals or status values from motors or valves, or, in particular, they can be actual or target values from control loops. Furthermore, signals can also be status values available in the system and assigned discontinuously to specific points in time, such as laboratory values for paper drums or the running time of fabric coverings. And the signals can also be image data or data derived from images, or acoustic data. These signal values can also originate from multiple process control systems (DCS) and multiple quality control systems within the production system.By using signals not only from a DCS but also from a QCS, quality changes are also detected, which significantly improves segmentation into different categories and model building. For example, quality changes can indicate changes in raw material quality or faulty production settings that increase the risk of disruption.
[0013] The data can be provided, for example, by reading in, transferring, or storing the data.
[0014] The paper machine, along with its associated stock preparation, raw material preparation and feeding, and existing downstream processing facilities, as well as any auxiliary equipment and laboratories, or a part thereof, is considered a production system.
[0015] The initial data set can encompass all or most of the available signals from the production system. The efficiency of the process can be improved by using a process model to reduce the number of signals included in this initial data set. This can be achieved, for example, by evaluating correlations between signals and / or by using only the signals relevant to the production disturbances being categorized.
[0016] Historical signal values, as used here, refer to signal values that describe production over a specific period in the past, for example, several weeks or months. The data can originate from a data storage system within the production system or be recorded intentionally over a specific period.
[0017] Selecting the points in time at which a production disruption, particularly a web break, occurred can be done manually or by inputting a list of such points. Preferably, this selection can be automated by using one or more signals from the initial data set to detect a disruption. For example, a disruption sensor, particularly a web break sensor or a quality sensor, can be used for this purpose. Signals indirectly linked to the production disruption can also be used, such as the rotational speeds of certain drives or pumps, or spray nozzles on a pulper that are activated upon web break.
[0018] The assignment of a characteristic numerical value to a signal change can be performed using an algorithm, particularly with the support of artificial intelligence methods. For assessing the relevance of a signal change, different time intervals prior to the disturbance can be used for various signals. In particular, the assignment considers not only the magnitude of the change, but also its dynamics, gradient, and other parameters derivable from the time course. Additionally, a weighting factor can be applied to each signal when assigning characteristic numerical values. These weighting factors can preferably be generated from a process model that takes into account the correlations between different signals. Each characteristic numerical value represents the change of a specific signal and is uniquely assigned to it.
[0019] The series of characteristic numerical values each describe the change in state of the production system before a disturbance. These series can be a type of vector, or ordered lists or matrices.
[0020] Various agreement criteria can be used to determine the substantial agreement of characteristic numerical values in series (R1...Rn). In particular, mathematical procedures for calculating the distance between the series are employed. Such mathematical procedures are generally known. If the distance is small, i.e., if there is substantial agreement, the series are assigned to the same category. Similar changes in certain signals and / or changes in the same signals thus lead to series being grouped into one category. That is, disturbances are grouped together in which similar changes of state occurred before the disturbance.
[0021] Determining the similarity between series of characteristic numerical values can preferably be done using artificial intelligence methods for pattern recognition or segmentation.
[0022] Optionally, the number of categories into which the segmentation should occur can be specified. For example, the matching criterion can be adjusted to create the desired number of categories.
[0023] It is advantageous to use at least 5 different, preferably at least 10 different, categories of production disruptions, especially web breaks. This allows for a sufficiently nuanced differentiation between the various causes of a production disruption. With too few categories, the corresponding data sets would still be so dissimilar that a reliable predictive model could not be formed. Alternatively, too many data sets would be relegated to the residual set and thus excluded from the analysis.
[0024] Furthermore, it is advantageous to use no more than 20, preferably no more than 30, different categories of production disruptions, especially web breaks. This avoids creating categories that are too small with too few associated series, the evaluation of which would result in less reliable predictive models.
[0025] To reduce computational effort and improve the quality of predictive models, when summarizing characteristic numerical values into series and / or determining the similarity between series, only a subset of the characteristic numerical values, i.e., only a subset of the underlying signals, can be used. Which characteristic numerical values, and thus which signals, are used can be defined via a process model, as described above.
[0026] Preferably, for each category (K1...Km) of production disruption, one or more corrective actions (M1...Mq) are assigned and stored in the procedure, with each action representing a change in production settings such that a reduction in the corresponding disruption risk of that category is expected. Thus, not only is a category represented, but also one or more corrective actions to reduce the risk of occurrence of this type of production disruption.
[0027] Any display or storage device can be used as the output device. Output can also be displayed in the DCS and / or QCS. The output can be displayed directly or stored for later viewing or use. It is advantageous to also output the suggested measures for reducing the risk of occurrence that belong to each category.
[0028] Predictive models are created using mathematical methods. The target value for each of these models is a so-called significator. This significator should be identifiable for a second dataset of signal values, represented as numerical sequences from a current production run, using the respective predictive model. The significator represents a value that describes the risk of disruption in the corresponding category. For example, the significator can be a single numerical value, specifically between 0 and 1. The closer it is to 1, the more likely a production disruption, particularly a track break, becomes in that corresponding category.
[0029] Predictive models are preferably created using machine learning methods, especially neural networks.
[0030] Other production disruptions that can be categorized using this method, besides web breaks, include quality deviations, meaning that certain quality parameters fall outside the specified range. These can include variations in basis weight, moisture content, and optical or paper-related quality parameters. Here, too, proper production is interrupted if the corresponding quality parameters are outside the permissible limits for the product.
[0031] According to the invention, the method for monitoring the production of a paper machine according to claim 8 is characterized by the following steps being carried out: 8.1 Providing categories (K1...Km), generated using a method according to any one of claims 1 to 6, each with an associated predictive model (P1...Pm). 8.2 Providing a second data set (D2) of current signal values of numerous different signals (S1...Sx) from a production system (10) of a paper machine as time series. 8.3 Determining one significator (X1...Xm) for each of the categories (K1...Km) by evaluating the second data set (D2) with the different predictive models (P1...Pm) for the different categories, each significator (X1...Xm) describing a category-specific risk of malfunction. 8.4 Output (A1, A2, A3, A4) at least those signifiers (X1...Xm) that reach or exceed a corresponding threshold (G1...Gm) to an end device (24) and / or in the DCS (20) and / or in the QCS (21). 8.5 Repeat steps 8.2 to 8.4. At another time, to continuously monitor production. This repetition can preferably be repeated at short intervals.
[0032] A significant advantage of the inventive embodiment of the method is that the current, ongoing production is monitored and impending production disruptions are detected in a timely manner and clearly assigned to a specific category. This is made possible by the category-specific output of the respective disruption risk.
[0033] The steps do not necessarily have to be performed in the specified order; they can also be carried out in a different sequence. Furthermore, individual steps can be executed in parallel. The characteristics for executing individual steps and for the data of the categorization process, already described above, can also be advantageously applied to the steps mentioned here without needing to be explicitly listed again. Similarly, the execution characteristics described below can also be advantageously used for the categorization process.
[0034] Current signal values refer to data from current production. This includes not only the actual values at the present time, but also signal values from production that occurred some time ago, i.e., from a period of several hours or days prior to the current time. Only in this way can the signal values be recorded as time series and thus changes in the system's state be evaluated.
[0035] Preferably, a process model can be used to reduce the number of required signals, as was done for the first data set. Particularly preferably, the second data set contains the signal values of the same signals that were also included in the first data set. Most preferably, the second data set contains the signal values of the signals that are used in at least one predictive model of one of the established categories.
[0036] The second dataset can be analyzed using the respective predictive models, employing machine learning methods. The target value of each analysis is a significator for each category. This significator indicates the current disruption risk for a specific category, based on the second dataset containing signal values from current production. The higher the corresponding current disruption risk, the more likely a production disruption is to occur in that category.
[0037] In particular, the significator is exactly one value. Advantageously, the significators are normalized to take values between 0 and 1. The closer the significator is to 1, the higher the risk of interference in this category.
[0038] A threshold can be set for each significator. Different thresholds can be defined for different categories. When a significator reaches or exceeds the threshold of its corresponding category, the significator is displayed. This indicates an increased risk of disruption in that category. All significators can also be displayed, and those that have reached their threshold can be highlighted. Furthermore, the temporal development of the significators can be displayed, allowing the evolution of disruption risks for the various categories to be monitored in relation to current production.
[0039] It is particularly advantageous if, in steps 8.2 and 8.3, not only the time series of the signal values, but also signal changes in a time interval and / or the gradient and / or other derived quantities of the signal value time series are used.
[0040] Furthermore, it is advantageous if, in step 8.4, additional measures are output that are assigned to the category to which a significator has reached or exceeded the threshold.
[0041] An additional advantage is that the operator receives suggestions for modifying ongoing production directly tailored to the specific situation. For each category of production disruption, the current risk of disruption and suitable corrective measures are provided separately. In this way, a large number of production disruptions, especially web breaks, can be avoided by making timely and, above all, targeted adjustments to production settings.
[0042] In order to improve the process and adapt it to changing production conditions, it is advantageous to provide for one or more repetitions of step 8.1 in order to update the categories and / or the predictive models and / or the measures after a certain production time.
[0043] Furthermore, the problem is solved by a computer program product for executing one of the methods according to the invention on a computer unit. This computer unit can be an independent unit or a separate computer; it can also be part of a process control system (DCS) or a quality control system (QCS).
[0044] Further advantageous features of the invention are explained using exemplary embodiments with reference to the drawings. Fig.1 Schematic representation of a production system with a paper machine and associated control and regulation systems as well as with a computer unit for carrying out a method according to the invention. Fig.2a / b Symbolic representation of a method according to the invention for classifying production defects Fig.3 Symbolic representation of a method according to the invention for avoiding production disruptions
[0045] The Fig.1 Figure 10 shows production system 10 with paper machine 1, stock preparation 2, and converting 3. Paper machine 1 is a machine for producing and / or processing fibrous webs, in particular a machine for producing paper or board webs, for example, a so-called board machine, or a pulp machine, a coating machine, or a paper converting machine. Converting 3 can be a reeling machine, a rewinding machine, a slitting machine, or a packaging machine. It can also be a coating machine or paper converting machine if the paper machine is a machine for producing paper or board webs. Stock preparation 2, paper machine 1, and converting 3 can also consist of several subsystems.Furthermore, the production system 10 includes the auxiliary units 4,5 and, if applicable, laboratories, as well as at least the process control system (DCS) 20 and at least the quality control system (QCS) 21.
[0046] The various signals Sx of the production system 10 are received, processed and controlled in the process control system (DCS) 20 and in the quality control system (QCS) 21 in order to control the production system as is known in the state of the art.
[0047] To execute the procedure for categorizing production disruptions, the first data set D1, comprising signal values from various signals Sx as time series Sx(t), is transferred to a computer unit 22. Optionally, these signal values can be stored beforehand on a data storage device 23 and retrieved as needed. The first data set D1 comprises historical signal values describing a period in which at least several production disruptions occurred. Typically, signal values from a period of several weeks or several months are used.
[0048] Computer Unit 22 can be a separate computer or part of the DCS or QCS.
[0049] The process for categorizing production disruptions is executed on computer unit 22, and the results are output. These results include, in particular, the identified categories K1...Km and / or the generated predictive models P1...Pm for the various categories. The output can be sent to different devices. For example, output A1 can be transmitted to DCS 20, output A2 to QCS, and / or output A3 to terminal 24. Terminal 24 can be a monitor or another computer, such as a tablet. Terminal 24 can then send further output A4 to QCS or to DCS.
[0050] When applying the procedure for preventing production disruptions, a second data set D2 containing current signal values of the signals Sx is transmitted as a time series Sx(t) to computer unit 22. Computer unit 22 applies the predictive models P1...Pm and evaluates the second data set D2 accordingly. The results of the procedure are then output as A1, A2, A3, and / or A4, as described previously. The results of this procedure are the significators X1...Xm, each indicating the category-specific disruption risk.
[0051] In Fig.2a und 2b The process of a categorization method according to the invention is shown schematically. Fig.2a This describes the formation of the series R1...Rn from the characteristic numerical values ZW1...ZWq. The first data set D1 comprises the signal values of numerous signals Sx of production system 10 as time series Sx(t). Various disturbance times t1, t2, t3, ... tn are selected and / or automatically determined. For each of these selected disturbance times, the signal changes ΔSx(tn) that occurred in a time interval before the disturbance time tn are determined for the various signals. Different time intervals can be used for different signals.
[0052] For example, shorter time intervals may be relevant for drive data, while time intervals of several hours may be relevant for raw material data, and even several days may be used for covering data, such as felt age. Furthermore, a process model can be used to select signals Sx that are provided in the initial data set D1 and / or for which the signal changes ΔSx(tn) are determined.
[0053] Each signal change ΔSx(tn) is assigned characteristic numerical values ZW and grouped into series, specifically vectors. Each numerical value ZW1...ZWq corresponds to a specific signal S1...Sx. And each series corresponds to a specific disturbance time t1...tn.
[0054] For the perturbation time t1, the series R1 is formed, here consisting of ZW2, ZW1, ZW3, ZWx..., for the perturbation time t2 the series R2, and so on up to the series Rn for the perturbation time tn. The series can be formed, in particular, as a type of vector, or as ordered series or matrices.
[0055] To determine the characteristic numerical value ZWx for a signal Sx, the change in the signal value over a time interval is used. This takes into account not only the absolute change, but also the dynamics of the change, the gradient, and other mathematical parameters of the time course. Additionally, a weighting factor can be applied to different signals. This allows, for example, the consideration of known relationships.
[0056] Fig. 2b This section describes the segmentation of series into different categories and the creation of the respective predictive models. For this purpose, the series R1...Rn are examined for similarity. Similarity between series is assumed if there is a substantial correspondence in the characteristic numerical values ZW1...ZWq contained within them. This substantial correspondence can be determined, in particular, by mathematically calculating the distance. If the distance is less than a defined value, the series are assigned to the same category. The distance calculation is performed using established mathematical methods. Furthermore, mathematical methods for pattern recognition or segmentation can be used to segment the series into different categories K1...Km.
[0057] Series identified as similar (R1...Rn) are assigned to the same category (K1...Km). The number of categories (K1...Km) can be predefined or generated by the evaluation algorithm. Series that are not similar to other series are assigned a residual set (L) and are not considered further in the evaluation. This could be due to production disruptions caused by external factors not reflected in the signal values. For example, if contaminants detach from the paper machine and fall into the paper web, causing a break.
[0058] Within each category, a predictive model P1...Pm is then created by evaluating the series assigned to that category. Artificial intelligence methods, particularly machine learning techniques such as neural networks, can be used for this purpose. A separate predictive model P1...Pm is created for each category K1...Km; this is a significant advantage of the method and offers higher reliability. The predictive models P1...Pm are designed to output a significator X1...Xm for a second dataset D2, indicating the category-specific disruption risk. In other words, they indicate the likelihood of a production disruption occurring in the respective category, as represented by the second dataset D2.
[0059] For example, different types of web breaks can be identified as categories. The segmentation can then differentiate web breaks that are heralded by fluctuations in stock preparation, by noticeable changes in the paper machine drive data, by changes in the moisture profile, or by other specific anomalies in the data.
[0060] The schematic representation in Fig.3 This shows the process for preventing production disruptions. The second data set, D2, comprises numerous signal values as time series Sx(t) from current production operations. If necessary, the number of signals included can be reduced by using a process model to select only the relevant signals for the types of production disruption under consideration.
[0061] Using the various predictive models P1...Pm, these time series are evaluated for each of the different categories. This can be done sequentially or in parallel for the categories.
[0062] Each predictive model P1...Px provides a significator X1...Xm that describes the risk of disruption for the respective category. Specifically, the significator is a single value and preferably lies between 0 and 1. The closer it is to 1, the higher the risk of a production disruption occurring in that category. Of course, other normalizations for the significator are also conceivable.
[0063] One of the particular advantages of the method according to the invention is that a separate significator is output for each category. This allows for very good monitoring of the production process and better conclusions to be drawn in the event of impending malfunctions.
[0064] Each significator X1...Xm is now compared to its corresponding threshold G1...Gm. If the threshold G1...Gm is reached or exceeded, there is a high risk of the corresponding production disruption occurring. At least the significators that exceed the respective threshold are output. Alternatively, all significators X1...Xm can be output, and those that have reached the threshold can be specially marked or highlighted.
[0065] Each category of production disruptions can be assigned one or more measures to reduce the disruption risk. For example, measures M1 and M2 are assigned to category K1, and measure M2 to category K2. In production, these are suggested changes to the production settings. For instance, the web tension can be increased if a category with fluttering web edges has been identified. Or, a change to the dewatering settings in the forming section can be assigned if a category with poor raw material dewaterability has been identified.
[0066] The measures M1...Mq assigned to the respective category are preferably also output, at least for the significators that have reached the threshold. This allows the operator or the process control system to initiate category-specific corrective actions in a timely and targeted manner, thus more reliably preventing impending production disruptions. Bezugszeichenliste
[0067] 1 Paper machine 2 Stock preparation 3 Further processing 4, 5 Auxiliary units 10 Production system 20 Process control system (DCS) 21 Quality control system (QCS) 22 Computer unit 23 Data storage 24 Terminal device A1, A2, A3, A4 edition D1 first data set D2 second data set S1, S2... Sx signal values ΔSx... z signal changes t1... n disturbance times ZWa, b, c... q characteristic numerical values R1,R2...RnSeries P1,P2...PmPredictive Models K1, K2...m Categories LRest quantity M1...q Measures X1, X2...XmSignifikatoren G1,G2...GmSchwellwerte
Claims
1. A method for categorising production faults on a paper machine (1), in particular web breaks, wherein the following steps are carried out: 1.1 Providing a first data set (D1) of historical signal values of numerous different signals (S1...Sx) from a production system (10) of a paper machine as time series (S1(t)...Sx(t)), wherein the signals originate at least in part from a process control system (DCS) (20), in particular from a drive control system, and at least partly from a quality control system (QCS) (21) of the production system (10). 1.2.1 Selecting a fault occurrence time (t1...tn) at which a production fault, in particular a web break, has occurred; for a plurality of signals from the first data set (D1), in particular for each signal, determining the change (Δ Sx...Δ Sz) in the respective signal value over a time interval prior to the selected fault time; and assigning a characteristic numerical value (ZW1...ZWx) to each signal change, wherein this numerical value (ZW1...ZWx) is higher the more significant the signal change is, and wherein each numerical value (ZW1...ZWx) is associated with a signal (S1...Sx). 1.2.2 Summarising the individual characteristic numerical values (ZW1...ZWx) for the selected disturbance time into a series (R1...Rn). 1.3 Repeating steps 1.2.1 and 1.2.2 several times, in each case for further selected disturbance times (t 1...tn), so that further series (R1...Rn) of characteristic numerical values (ZW1...ZWx), assigned to the respective disturbance times, are formed. 1.4 segmenting the series (R1...Rn) thus obtained into different categories (K1...Km), by assigning similar sequences (R1...Rn) to the same categories (K1...Km) respectively, whereby sequences (R1...Rn) are regarded as similar if they show substantial agreement in the characteristic numerical values (ZW1...ZWx); each category (K1...Km) represents a type of production fault, in particular a type of web break. 1.5 Output (A1, A2, A3, A4) of the formed categories (K1...Km) and / or the series (R1...Rn) assigned to the respective category to a terminal device (24) and / or in the DCS and / or in the QCS. 1.6 Forming a respective predictive model (P1...Pm) for each of the categories (K1...Km) by evaluating the series (R1...Rn) assigned to the respective category (K1...Km), whereby a separate predictive model (P1...Pm) is formed, and in such a way that the predictive models (P1...Pm) are suitable for evaluating a second data set (D2) of signal values as time series and for specifying, for each respective category (K1...Km), a signifier that describes a category-specific fault risk.
2. Method according to claim 1 wherein one or more measures (M1...Mq) are assigned to and stored for each category (K1...Km) of production disruption, each of which represents a change to the production settings, such that a reduction in the corresponding risk of disruption for that category is to be expected.
3. A method according to claim 2 wherein the measures (M1...Mq) for each category (K1...Km) are output (A1, A2, A3, A4) to a terminal (24) and / or in the DCS and / or QCS.
4. A method according to one of the preceding claims wherein at least 5 different, preferably at least 10 different categories (K1...Km) of production faults, in particular web breaks, are used, and / or that at most 20 different categories (K1...Km) of production faults, in particular web breaks, are used.
5. A method according to one of the preceding claims wherein the selection of a fault time (t1...tn) at which a production fault, in particular a web break, has occurred is carried out automatically by using a single or several signals from the first data set (D1) as an indicator of a production fault.
6. A method according to one of the preceding claims wherein the predictive models (P1...Pm) are formed using machine learning methods, in particular neural networks.
7. A computer program product for executing a method according to one of the preceding claims on a computer unit (20, 21, 22).
8. A method for monitoring the production of a paper machine (1), in particular for web breaks, using a categorisation (K1...Km) which was created using a method according to any of claims 1 to 6, wherein the following steps are performed: 8.1 Providing categories (K1...Km), formed using a method according to any one of claims 1 to 6, each with an associated predictive model (P1...Pm). 8.2 Providing a second data set (D2) of current signal values for numerous different signals (S1...Sx) from a production system (10) of a paper machine in the form of time series. 8.3 Determining a respective signifier (X1...Xm) for each of the categories (K1...Km) by evaluating the second data set (D2) in each case using the various predictive models (P1...Pm) for the various categories, wherein the signifier (X1...Xm) describes the category-specific risk of failure. 8.4 Outputting (A1, A2, A3, A4) at least those signifiers (X1...Xm) which reach or exceed a corresponding threshold value (G1...Gm) to a terminal device (24) and / or in the DCS (20) and / or in the QCS (21). 8.5 Repeating steps 8.2 to 8.4 to continuously monitor production.
9. A method according to claim 8, using a categorisation created by means of a method according to any one of claims 2 to 6, wherein the signifiers are normalised such that they assume values between 0 and 1.
10. A method according to claim 8 or 9 wherein in steps 8.2 and 8.3, not only the time series of the signal values but also the signal changes over a time interval and / or the gradient and / or other derived quantities of the signal value time series are used.
11. A method according to any one of claims 8 to 10, using a categorisation which was created by means of a method according to claim 5, wherein in step 8.4, additional measures (M1...Mq) are output (A1, A2, A3, A4) that are assigned to the category (K1...Km) for which a signifier (X1...Xm) has reached or exceeded the threshold value (G1...Gm).
12. A method according to any one of claims 8 to 11 wherein one or more repetitions of step 8.1 are provided in order to update the categories (K1...Km) and / or the predictive models (P1...Pm) and / or the measures (M1...Mq) after a certain production time.
13. A computer program product for executing a method according to any one of claims 8 to 11 on a computer unit (20, 21, 22).