Adaptive learning, automatic labeling method and system for predicting and diagnosing paper breaks in a paper machine

Abstract

Embodiments of the present invention disclose a method and system (103) for labelling normal and abnormal regions in data related to a paper machine (101) for web break prediction and labelling various parameters for root cause analysis using machine learning models. Further, the present invention relates to using the generated labels to train machine learning models to predict web breaks and root causes of web breaks. Thereafter, the present invention includes using machine learning models in real-time to predict breaks in the web, analyze root causes of breaks in the web and estimate break times. The proposed automated data labelling framework facilitates autonomous model improvement of deployed models, transfer learning, adaptive learning of reduced parameters and automated feasibility studies.

Description

Technical Field

[0001] This invention generally relates to a framework for predicting faults in equipment. More specifically, the invention relates to marking, predicting, and diagnosing breaks in the paper web of a papermaking machine. Background Technology

[0002] Paper is consumed in various forms in our daily lives, such as packaging paper, writing paper, printing paper, cardboard, corrugated board, tissue paper, books, magazines, documents, and newspapers. Recently, paper has seen significant growth in the packaging sectors of both industrial and consumer goods industries. The papermaking process involves a paper machine, an industrial machine that converts large quantities of pulp into paper at high speed. Pulp parameters or properties differ for different grades (e.g., the base weight of tissue paper differs from that of cardboard or writing paper). Pulp is fed into the paper machine, where it is mixed with water. The paper machine produces different grades of paper depending on the application using the relevant pulp.

[0003] A paper machine involves multiple parts in the papermaking process. These parts include the headbox, forming section, press section, drying section, calendering section, and winding section. Modern paper machines operate at 5000 ft / min. In paper machines, a common problem is a break called web break (a break in the paper web), which causes production downtime. Therefore, predicting web break, its expected timing, and possible root causes will help operators avoid breakage in such situations or in the future.

[0004] Various solutions have been proposed to predict paper web breakage and analyze its root causes. Few techniques utilize neural networks for predicting breakage and decision trees for root cause analysis. Existing techniques have limitations in terms of the labeled data required for predicting breakage and analyzing root causes. In existing techniques, data is manually labeled, which becomes cumbersome and impractical as the number of parameters in a paper mill increases. Furthermore, unsupervised models used with unlabeled data to predict breakage do not yield robust predictions. Therefore, there is a need to label data effectively to reduce manual effort and accurately detect transitions leading to breakage. Summary of the Invention

[0005] In one embodiment, the present invention relates to a method for labeling parameters associated with a papermaking machine to predict breakage in the paper web caused by the papermaking machine. The parameters include multiple process parameters and multiple operational parameters. In one embodiment, the parameters are not limited to individual process and operational parameters, but the invention can utilize any signals measured in the papermaking workshop. In one embodiment, simulated parameters associated with the papermaking machine are provided to multiple machine learning models, the simulated parameters including normal patterns and anomalous patterns causing breakage in the paper web. Known techniques can be used to generate the simulated parameters. In one embodiment, historical parameters of the papermaking workshop are used to generate the simulated parameters. Furthermore, the multiple machine learning models are configured to label the simulated parameters as normal and anomalous patterns. In one embodiment, the anomalous pattern is close to the timestamp of breakage in the paper web. Additionally, outputs from each of the multiple machine learning models are received, where the outputs indicate labels including normal and anomalous patterns. Furthermore, a model is selected from the multiple machine learning models based on one or more performance metrics and the outputs of the multiple machine learning models, and one or more model parameters of the selected model are stored in the memory of a computing unit. Subsequently, historical parameters related to the paper machine, along with several details of the selected model, are provided to the automatic labeling machine to label the historical parameters as normal and abnormal modes, including at least one of these modes. The labels generated by the automatic labeling machine are stored as labeling data in a database, where this data is used to predict breakage in the paper web and the root causes of such breakage in real time.

[0006] In one embodiment, simulation parameters and historical parameters are received in multiple batches. Each batch includes multiple process parameters and multiple operational parameters, which are simulated or measured between the time when the paper machine is restarted after the paper web breaks and the time when the paper web subsequently breaks in the paper machine.

[0007] In one embodiment, simulation parameters and historical parameters with anomalous patterns are labeled as the root cause of breakage in the paper web, wherein the root cause of breakage in the paper web is included in the labeled data.

[0008] In one embodiment, simulation parameters and historical parameters with anomalous patterns are tagged as estimated paper web breakage times, wherein the estimated paper web breakage times are included in the tagged data.

[0009] In one embodiment, each batch is labeled as including a normal mode and an abnormal mode, based on independent analysis of each batch or by providing one of similar batches to an automated labeling machine.

[0010] In one embodiment, the present invention relates to training multiple machine learning models to predict breaks in paper webs. Training the multiple machine learning models includes providing labeled data, comprising normal and abnormal patterns including parameters related to a paper machine, to the multiple machine learning models. In one embodiment, timestamps are associated with outliers, and the timestamps of the outliers are close to the timestamps of breaks in the paper web. The labeled data is generated using an automated labeling machine with simulated parameters and historical parameters related to the paper machine. Furthermore, the multiple machine learning models are configured to detect patterns in the parameters and determine the detected patterns as at least one of normal and abnormal patterns. Furthermore, the multiple machine learning models are configured to generate expected patterns for each parameter based on the determination, wherein each expected pattern includes at least one of normal and abnormal patterns. Additionally, the output from each of the multiple machine learning models is received, wherein the output represents a prediction of breaks in the paper web based on a comparison of the labeled data with the corresponding expected pattern. Subsequently, a machine learning model is selected from the multiple machine learning models based on the outputs of the multiple machine learning models.

[0011] In one embodiment, the present invention relates to a method for predicting breakage in the paper web within a papermaking machine. Multiple sensors are used to monitor parameters associated with the papermaking machine, said parameters including multiple process parameters and multiple operational parameters. During the papermaking process, a paper web is fed / formed in the papermaking machine. The method includes receiving parameters from the multiple sensors. Furthermore, a pattern of change for each parameter over time is determined. Additionally, each of the determined patterns is compared to a corresponding expected pattern. Breakage in the paper web is then predicted based on the comparison.

[0012] In one embodiment, multiple models are also trained to estimate the time of paper web breakage and to determine the root cause of the breakage based on anomalies in the parameters and labeled data. Furthermore, multiple models are trained to generate correlations between anomalies and the estimated time of paper web breakage, as well as correlations between anomalies and the root cause of the breakage in the paper web.

[0013] In one embodiment, the selected model identifies anomalous patterns in parameters received from one or more sensors based on a comparison of each of the determined patterns with a corresponding expected pattern. Furthermore, the selected model determines the root cause of the paper web breakage and the estimated paper web breakage time based on the correlation between the anomalous patterns and the root cause of the breakage, as well as the correlation between the anomalous patterns and the estimated paper web breakage time.

[0014] In one embodiment, parameters are received after each tear in the paper web. Furthermore, these parameters are provided to an automatic labeling machine to generate labels that include both normal and abnormal patterns. Additionally, the generated labels are stored as label data in a database.

[0015] In one embodiment, tagged data stored in a database is fed to a machine learning model as feedback at predetermined time intervals. Furthermore, one or more features are identified from the tagged data that cause tears in the paper web. The machine learning model is then trained to adapt to the one or more features and configured to predict tears in the paper web based on those features.

[0016] In one embodiment, a prescriptive analysis is performed on multiple processing parameters and multiple operating parameters, as well as at least one parameter identified as the root cause of breakage in the paper web, to determine one or more actions to be performed to restore the operation of the paper machine.

[0017] In one embodiment, one or more model parameters of the machine learning model are a new machine learning model used to predict breakage in the paper web, determine the root cause of breakage in the paper web, and estimate the breakage time of the paper web in a new papermaking machine.

[0018] Systems of varying ranges are described herein. In addition to the aspects and advantages described in this overview, other aspects and advantages will become apparent from the accompanying drawings and the following detailed description. Attached Figure Description

[0019] The subject matter of the invention will be explained in more detail below with reference to preferred exemplary embodiments shown in the accompanying drawings, in which:

[0020] Figure 1 A framework for marking and predicting breakage in paper webs in a papermaking machine is shown according to some embodiments of the invention;

[0021] Figure 2 An exemplary flowchart for labeling parameters related to a papermaking machine is shown according to some embodiments of the present invention;

[0022] Figure 3A The training and inference phases of a machine learning model for labeling parameters related to a papermaking machine, according to some embodiments of the present invention, are illustrated.

[0023] Figure 3B Exemplary graphs of simulation parameters according to some embodiments of the present invention are shown;

[0024] Figure 3C A graph illustrating the scores of different time values ​​combined by different ML models according to some embodiments of the present invention is shown;

[0025] Figure 4A and Figure 4B An exemplary graph illustrating markers indicating abnormal patterns in parameters related to a papermaking machine is shown according to some embodiments of the present invention;

[0026] Figure 5 Exemplary graphs are shown for selecting parameters that contribute to breakage in the paper web, according to some embodiments of the present invention.

[0027] Figure 6A and Figure 6B An exemplary graph is shown, according to some embodiments of the present invention, representing anomalous patterns in reduced parameters associated with a paper machine;

[0028] Figure 7A and Figure 7B A flowchart for training and inferring breakage prediction in paper web is shown according to some embodiments of the present invention;

[0029] Figure 8A The training and inference phases of a machine learning model for predicting breakage in paper webs, the root cause of breakage, and the estimated breakage time, according to some embodiments of the present invention, are illustrated.

[0030] Figure 8B An exemplary block diagram of adaptive learning for a selected ML model is shown according to some embodiments of the present invention;

[0031] Figure 9 Exemplary graphs illustrating marked anomaly patterns and predicted anomaly patterns are shown according to some embodiments of the present invention;

[0032] Figure 10 Exemplary graphs illustrating the use of semi-supervised techniques to predict anomalous patterns are shown according to some embodiments of the present invention;

[0033] Figure 11 This is a diagram illustrating the reconstruction error in the prediction of breakage in the paper web of a paper machine using semi-supervised technology according to some embodiments of this disclosure;

[0034] Figure 12A and Figure 12B An exemplary graph illustrating the learning of a machine learning model according to some embodiments of the present invention is shown;

[0035] Figure 13 Examples of methods for generating such data according to some embodiments of the present invention are shown. Figure 12A and Figure 12B The network diagram showing the training and testing results; and

[0036] Figure 14An exemplary block diagram is shown for transferring knowledge from one ML model to another associated ML model according to some embodiments of the present invention; Detailed Implementation

[0037] Embodiments of this invention disclose methods and systems for labeling parameters related to a paper machine using a machine learning model, wherein the labeling includes normal and abnormal modes. Furthermore, this invention relates to training a machine learning model to predict breakage in a paper machine using the labeling. Subsequently, this invention includes using the machine learning model in real time to predict breakage in the paper web, analyzing the root causes of breakage in the paper web, and estimating the breakage time. Breakage notifications are then sent to the operator for timely action.

[0038] Figure 1A framework (100) is shown for labeling parameters associated with a paper machine (101), predicting breakage of the paper web in the paper machine (101), analyzing the root causes of breakage in the paper web, and estimating the timing of breakage in the paper web. Typically, one or more paper machines (101) are used in a paper mill to manufacture paper. A paper machine includes multiple sections, and multiple sensors are used to monitor said multiple sections. Historical measurements from the sensors are stored in a history recorder and used for analysis, such as predicting breakage in the paper web, identifying the root causes of historical breakage in the paper web, etc. Typically, parameters associated with the paper machine (101) include multiple process parameters and multiple operational parameters. In one embodiment, parameters may include, but are not limited to, process parameters, drive parameters, raw material parameters, quality control parameters, mechanical parameters, and sensor parameters. In one embodiment, any parameter in the paper mill can be considered as contemplated by the parameters of this invention. In one embodiment, multiple process parameters may include, but are not limited to, pulp quality, pulp-to-water mixing ratio, wetting factor, etc. In one embodiment, multiple operational parameters may include, but are not limited to, drive speed, paper web drying time, airflow rate, drive temperature, manufacturing grade, etc. This invention is not limited to specific parameters of a paper machine, but can be modeled to consider multiple parameters of paper machines commonly used in a paper mill. The framework (100) includes a data processing structure (102) (e.g., the Lamda architecture) for providing data collected from the paper machine (101) for labeling normal and abnormal patterns and for predicting breaks in the paper web. The data processing structure (102) is used to process large amounts of data and provides access to both batch and stream processing. In a data processing architecture (102) such as the Lamda architecture, there are batch and stream layers. In one embodiment, data blocks (e.g., hourly, daily, weekly, and monthly data) are fed to the batch layer for processing. In another embodiment, data streams are fed to the stream layer (also known as the velocity layer). Typically, real-time parameters are fed to the stream layer. In this invention, historical parameters associated with the paper machine (101) are fed to the batch layer, while real-time parameters associated with the paper machine (101) are fed to the stream layer. In one embodiment, the batch layer corresponds to the offline analysis of parameters related to the paper machine (101) in this invention, while the flow layer corresponds to the online analysis of parameters related to the paper machine (101).

[0039] In one embodiment, a computing unit (103), such as a personal computer, laptop computer, server, or any other computing device, can be used to perform offline and online analysis. In one embodiment, the computing unit (103) may be a cloud environment and is connected to the paper mill via a network. The computing unit (103) includes one or more hardware processors and memory. Offline analysis is performed by the computing unit (103) to flag parameters related to the paper machine (101) as including normal and abnormal modes. Online analysis is performed using the flagged parameters to predict breaks in the paper web. Furthermore, the predicted breaks in the paper web are provided to a notification unit (104) to alert operators in the paper mill. In addition, the root cause of the break, the estimated break time, and one or more actions to be performed to avoid breaks in the paper web are notified on the notification unit (104). In one embodiment, the notification unit (104) may include, but is not limited to, a display unit, a speaker, an optical notification, and combinations thereof.

[0040] Figure 2 A flowchart (200) is shown for labeling parameters related to the paper machine (101). See Figures 3 and 4 for details. Figure 2 The method and steps.

[0041] In step (201), the computing unit (103) provides simulation parameters, including normal and abnormal modes, to multiple machine learning (ML) models. Now refer to... Figure 3A As shown in the training phase, simulation parameters are provided to multiple ML models (301a, 301b, ..., 301n). The simulation parameters are simulations of parameters associated with the paper machine (101). Typically, parameters associated with the paper machine include multiple anomalies. Moreover, due to the large datasets (typically thousands of parameters for the paper machine), it is not easy to identify the anomalies causing breakage in the paper web. Multiple anomalies can occur at any time during the operation of the paper machine (101). This invention proposes a robust labeling technique to identify specific anomalies causing breakage in the paper web. The simulation parameters are simulations of parameters causing breakage in the paper web. Therefore, the simulation parameters include normal and abnormal modes of the parameters. Figure 3B The image shows a graph illustrating the variation of exemplary simulation parameters (parameter magnitudes relative to time). Figure 3B As shown, the simulation parameters include multiple transitions (306a, 306b, 306c, 306d) and a break point (307). The break point (307) represents the timestamp when a break occurs in the paper web. Figure 3BAs shown, the parameters are normal after the transitions (306a, 306b, and 306c) have occurred. This is typically the case in a paper machine. The paper machine operates normally after several parameter transitions. However, after a specific transition (e.g., 306d), breakage occurs in the paper web. Observation reveals that the transition occurring near the breakage point (307) is the abnormality and the cause of the breakage in the paper web.

[0042] In one embodiment, the simulation parameters are divided into multiple batches. A batch may include multiple process parameters and multiple operational parameters that are simulated or measured between the time it takes for the paper machine to restart after a break in the paper web and the time it takes for the paper web to subsequently break again in the paper machine. For example, consider a first break in the paper web occurring at a first timepoint in the paper machine. The paper machine resumes operation at a second timepoint, and a second break in the paper web occurs at a third timepoint. A batch can be considered as parameters simulated or measured between the second and third timepoints.

[0043] Back Figure 2 In step (202), the computing unit (103) configures multiple ML models (301a, ..., 301n) to label simulation parameters. In one embodiment, the multiple ML models (301a, ..., 301n) can use various labeling techniques to label simulation parameters. In one embodiment, the multiple ML models (301a, ..., 301n) are configured to label anomalous patterns such that the anomalous patterns are close to the breakpoint (307). For example, transition (306d) is close to the breakpoint (307), while transitions (306a, 306b, and 306c) are far from the breakpoint (307). Therefore, the timestamp from the start of transition (306d) to the timestamp of the breakpoint (307) is labeled as an anomalous region including the anomalous pattern, and the remainder of this region is labeled as a normal region including the normal pattern. In one embodiment, the closer the anomalous pattern is to the beak point (305), the more accurate the output of the multiple ML models (301a, ..., 301n).

[0044] Now for reference Figure 4A , Figure 4AA graph showing the marked abnormal region (401) (parameter amplitude changes with time) is shown. As shown, the abnormal region (401) is determined by the break point (307). For example, the abnormal region could be the region fifteen minutes before the break point (307). In another example, the abnormal region could be the region 30 minutes before the break point (307). In yet another example, the abnormal region could be the region one hour before the break point (307). When multiple ML models (301a, ..., 301n) use the break point (307) as a reference to mark normal and abnormal regions, it can be considered that the multiple ML models (301a, ..., 301n) are using semi-supervised learning techniques. In one embodiment, normal and abnormal regions are marked for each batch to predict breakage in the paper web. In one embodiment, each parameter (each processing parameter and each operating parameter) is marked to analyze the root cause of breakage in the paper web. Figure 4B The parameters for marking normal and abnormal regions are shown. Figure 4B The marker (401) in the diagram indicates an abnormal region in each parameter. The abnormal region in each parameter is used to mark the root cause of breakage in the paper web. In one embodiment, parameters from multiple process parameters and multiple operating parameters are identified as root causes of breakage in the paper web. In one embodiment, a domain expert can identify different parameters as root causes based on abnormal patterns in the parameters. For example, an object falling onto the paper web may cause breakage. A tension sensor can measure the tension of the paper web, and the tension value may have an abnormal pattern of disturbance in the paper web caused by the falling object. A domain expert can observe the abnormal pattern in the tension value and identify the falling object as the root cause. In another example, the motor speed, as a measurement parameter, may have an abnormal pattern. An automatic labeling machine (304) can identify motor speeds with abnormal patterns causing breakage in the paper web from historical batches and mark the motor speed as the root cause. It is clear from the two examples above that the measured parameter may be the root cause of breakage in the paper web, or the measured parameter may exhibit an abnormal pattern due to external factors (e.g., falling objects), which may be the root cause of breakage in the paper web.

[0045] In one embodiment, the root cause of the breakage in the paper web is used to mark the estimated breakage time in the paper web. For example, an automatic labeling machine (304) can identify from historical batches that the paper web breaks within 5 minutes after the occurrence of an abnormal pattern in the motor speed. Based on a comparison with historical batches, the automatic labeling machine (304) can mark the estimated breakage time as 5 minutes for an input batch with an abnormal motor speed.

[0046] Refer again Figure 2In step (203), the computing unit (103) receives output from each of the multiple ML models (301a, ..., 301n). In one embodiment, each ML model outputs labels indicating normal and abnormal regions. In another embodiment, each ML model merges the normal and abnormal regions in the batch. Figure 3C An example graph illustrating the combined scores plotted by different ML models relative to time values ​​is shown. Figure 4A As shown, a graph can be plotted for each ML model. The anomalous region (401) in each graph can vary and can be closer to or further away from the breakpoint (307).

[0047] In step (204), the computation unit (103) selects an ML model from multiple ML models (301a, ..., 301n) based on the output and one or more performance metrics. Table 1 shows exemplary performance metrics and outputs for the three ML models.

[0048]

[0049] Table 1

[0050] In one embodiment, the performance metric “time_score_merge” indicates the proximity of an anomalous region (401) to the breakpoint (307). A zero value indicates an accurate prediction, while any negative value with an increased magnitude results in a normal region being predicted as an anomalous, and any positive value with an increased magnitude (loose boundary) results in an anomalous being predicted as normal, making the algorithm unreliable for labeling. For example, if the labeled data is used for single-class learning, an algorithm with a negative time score close to zero or exactly zero (tight boundary) would be selected. Similarly, the performance metrics “specificity_merge” and “sensitivity_merge” should be maximized. Likewise, many other custom performance metrics are considered to select an ML model from multiple ML models (301a, ..., 301n). Subsequently, one or more model parameters of the selected ML model (ML techniques and their hyperparameters may include, for example, approximate no of PCA components, distance measurement, percentage of neighbors, merge time, raw data, lagged data, noise filter window size, scaling criterion, percentage of contamination, etc.) are stored in processor-associated memory (303).

[0051] See again Figure 2In step (205), the calculation unit (103) provides one or more model parameters of the selected ML model and historical parameters associated with the paper machine (101) to the automatic labeling machine (304). In one embodiment, a merging time parameter is provided to the automatic labeling machine (304) along with one or more model parameters and historical parameters. The merging time parameter indicates the maximum time an anomaly is temporarily merged into a batch. In one embodiment, the merging time parameter is selected based on a performance metric. In another embodiment, one or more model hyperparameters and the merging time parameter are fine-tuned to tag historical parameters based on performance metrics. Figure 3C A graph showing the merging time of multiple ML models (301a, ..., 301n) is presented. In one embodiment, the automatic labeling machine (304) may be an ML model derived from multiple ML models (301a, ..., 301n). Historical parameters can be obtained from a historical recorder (not shown) associated with the paper mill. The historical recorder may store historical process parameters and operating parameters of the paper mill (101). In one embodiment, historical parameters may include multiple anomalies. Conventional techniques are not effective in detecting specific anomalies that cause breakage in the paper web. In this invention, one or more model parameters are used to mark normal and abnormal areas in the historical parameters. In one embodiment, historical parameters are provided as multiple batches. In one embodiment, the automatic labeling machine marks the entire batch as normal and abnormal areas, and also marks each process parameter and each operating parameter as normal and abnormal areas. The marked batches are stored as marked data in a database (305), and the marked data is used to predict breakage in the paper web. The marking of each process parameter and each operating parameter is used for root cause analysis and time estimation to predict breakage in the paper web.

[0052] In one embodiment, each batch parameter is associated with a grade, and each grade includes specific operating conditions. Examples of grades may include, but are not limited to, bond or writing grades, book grades, and text grades. In one embodiment, the grade may be determined based on the basis weight of the paper. In one embodiment, individual batches may be analyzed and labeled as normal and abnormal areas. For example, an automatic labeling machine (304) considers individual batch analysis of a given batch to label normal and abnormal areas. However, if the grade and operating conditions of a batch are known, it is compared with all other batches within the same grade and operating conditions to make the labeling more robust. For example, each batch may be associated with a grade, and each grade may be associated with one or more operating conditions. For example, when manufacturing security grade paper, the paper machine (101) may operate at 1500 m / min, while when manufacturing text grade paper, the paper machine (101) may operate at 2000 m / min. Similarly, each grade may be associated with different operating conditions and may also include different process conditions. Therefore, when a batch associated with a text grade and operated by the paper machine (101) at 2000 m / min is provided to the automatic labeling machine (304) for labeling, the automatic labeling machine (304) can compare the batch with similar historical batches belonging to the text grade and operating at 2000 m / min. Based on the comparison with similar historical batches, the automatic labeling machine (304) labels the batch.

[0053] Now for reference Figure 5 It shows an exemplary graph for selecting parameters that contribute to breakage in the paper web. In one embodiment, each parameter may be numbered, and as shown... Figure 5 As shown, a graph is plotted for each parameter relative to the number of breaks in the paper web caused by anomalous patterns in each parameter. A threshold is set, and one or more parameters exceeding the threshold can be selected as the most likely causes of breaks in the paper web. For example, 800 parameters can be recorded when a break occurs in the paper web. However, based on historical analysis, anomalous patterns are identified in the top 300 parameters when a break occurs. Therefore, the root cause of breaks in the paper web can be analyzed by considering only the top 300 parameters, and the timing of breaks in the paper web can be estimated. Simplifying the parameters reduces the dependence on computational resources.

[0054] Figure 6A It shows the use of Figure 5 The simplified parameters derived from the curve graph are used to label the curve graph of the batch, while Figure 6B It shows the use of Figure 6B The graph uses simplified parameters to represent the graph of each parameter in the batch. Figure 6A and Figure 6B The marker (601) indicates an abnormal pattern in a batch with reduced parameters.

[0055] Figure 7A and Figure 7B An online analysis of parameters related to the paper machine (101) is shown. (Reference) Figure 8A , Figure 8B and Figure 9 describe Figure 7A and Figure 7B . Figure 7A Training multiple ML models (801a, ..., 801n) to predict breakage in the paper web, predict the root causes of possible breakage in the paper web, and estimate the timing of breakage in the paper web are shown.

[0056] Reference Figure 7A In step (701), the computing unit (103) provides labeled data to multiple ML models (801a, ..., 801n) (e.g. Figure 8A (As shown). In one embodiment, the ML model (801a, ..., 801n) may differ from multiple ML models (301a, ..., 301n). The multiple ML models (801a, ..., 801n) are configured to predict breakage in the paper web, predict the root cause of possible breakage in the paper web, and estimate the time of breakage in the paper web. The labeled data includes labels for regions in batches of historical parameters and labels for each historical parameter, where anomalous regions are close to the breakage point (307).

[0057] See again Figure 7A In step (702), the computing unit (103) configures multiple ML models (801a, ..., 801n) to detect patterns and determine the detected patterns as either normal or abnormal patterns. The multiple ML models (801a, ..., 801n) detect patterns in the labeled data. A pattern can be a change in parameters. For example, the motor speed can gradually increase from 1000 rpm to 1100 rpm. Patterns of motor speed changes are detected. This pattern can be predicted as either a normal or abnormal pattern. For example, a change in motor speed between 1000 rpm and 1100 rpm can be determined as a normal pattern based on markers provided in the labeled data. In another example, a sudden increase in motor speed from 1000 rpm to 1100 rpm can be determined as an abnormal pattern. In one embodiment, markers in the labeled data are used as a reference for the multiple ML models (801a, ..., 801n) to determine whether a pattern is a normal or abnormal pattern. In one embodiment, a classification ML model can be used to determine whether a pattern with parameters of a specific state is a normal or abnormal pattern. In one embodiment, when parameters do not have a specific state (defined category) and parameter determination is based on parameter values, a regression ML model can be used to determine the pattern of these parameters as either a normal pattern or an abnormal pattern.

[0058] In step (703), the computing unit (103) configures multiple ML models (801a, ..., 801n) to generate a predicted pattern for each parameter. In one embodiment, the predicted pattern indicates a reference value for each parameter. In another embodiment, the normal pattern for each parameter is considered the predicted pattern. During real-time monitoring of the parameter, a break is predicted based on the deviation from the predicted pattern of the parameter, which includes the corresponding normal value. Because many anomalies in the parameter may not be known to label all outliers, multiple ML models (801a, ..., 801n) can be trained to identify normal values ​​as predicted patterns. In one embodiment, training ML models (801a, ..., 801n) to identify a single class (normal region) is called single-class learning. In one embodiment, the predicted pattern includes both normal and outlier values. A binary classifier can be used to classify the parameter into normal and outlier regions. In one embodiment, a binary classifier is used when sufficient outlier or anomalous data is used to label outlier regions.

[0059] In step (704), the computing unit (103) receives output from each of the plurality of ML models (801a, ..., 801n). The output of each of the plurality of ML models (801a, ..., 801n) represents a prediction of breakage in the paper web. Each of the plurality of ML models (801a, ..., 801n) compares the detected patterns with the expected patterns and predicts breakage in the paper web. For example, an alarm indicating breakage in the paper web is generated when the parameter patterns in the labeled data do not match the expected patterns. In one embodiment, the plurality of ML models (801a, ..., 801n) are also configured to determine the root cause of breakage in the paper web based on the parameters and the anomalous patterns in the labeled data and to estimate the timing of breakage in the paper web. Once all parameters are determined to be either a normal pattern or an anomalous pattern, each of the plurality of ML models (801a, ..., 801n) determines the root cause of breakage in the paper web. Multiple ML models (801a, ..., 801n) use labeled data to determine root causes during training. Similarly, multiple ML models (801a, ..., 801n) use labeled data to estimate the paper web breaking time during training. During training, because the labeled data includes labels for historical parameters (typically a large dataset), multiple ML models (801a, ..., 801n) can generate correlations between parameters with anomalous patterns and the identified root causes. Likewise, multiple ML models (801a, ..., 801n) can generate correlations between parameters with anomalous patterns and the estimated paper web breaking time. For example, when an object falls onto the paper web, the tension values ​​of the paper web include anomalous patterns. During training, when an object falls onto the paper web, when sufficient tension values ​​with anomalous patterns are provided, multiple ML models (801a, ..., 801n) can identify anomalous patterns in the tension parameters and correlate them with the root cause when such an anomalous pattern is identified. Considering the example above, when the tension parameter exhibits an abnormal pattern, the break point is 3 minutes away from the occurrence of this abnormal pattern. During training, when provided with sufficient tension values ​​exhibiting at least one of the normal and abnormal patterns, multiple ML models (801a, ..., 801n) can identify the normal or abnormal patterns in the tension parameter and estimate that the paper web will break within 3 minutes.

[0060] In step (705), the computing unit (103) selects an ML model (e.g., 801a) from the multiple ML models (801a, ..., 801n) based on their outputs. In one embodiment, a model selector (802), which may be a module implemented by the processor of the computing unit (103), may select an ML model (e.g., 801a) based on the performance of the multiple ML models (801a, ..., 801n). For example, an ML model (801a) is selected when it meets one or more performance metrics associated with the multiple ML models (801a, ..., 801n). Thereafter, the model selector (802) stores the selected model (e.g., 801a) in memory (303). When multiple ML models (801a, ..., 801n) are supervised models, metrics based on the confusion matrix are used (e.g., accuracy, precision, recall, AUC-ROC, F1 score, etc.). If the model is semi-supervised (single-class learning), mean squared error (MSE), R^2, etc., can be used. For example, if an encoder-decoder mechanism is used, MSE can be used as a performance metric.

[0061] In one embodiment, the training of multiple ML models (801a, ..., 801n) can be performed using supervised or semi-supervised techniques. In one embodiment, supervised techniques can be used when the labeled data includes labels based on rank and operating conditions, while semi-supervised techniques or single-class learning can be used when the labeled data includes labels based on independent batches.

[0062] Now for reference Figure 7B This demonstrates the inference performed in real time by a selected ML model (e.g., 801a). In step (706), the computing unit (103) receives parameters related to the paper machine (101) from multiple sensors in real time. In one embodiment, the computing unit (103) may receive the parameters via a data processing architecture (102). The data processing architecture (102) streams the real-time parameters to the computing unit (103), where the received parameters are time-series data. The data processing architecture (102) may preprocess the parameters to reduce noise, increase the strength of the parameters, etc.

[0063] In step (707), the calculation unit (103) determines the pattern of change of each parameter over time. In one embodiment, the calculation unit (103) can sample the parameters at high speed to determine minute changes in the parameters.

[0064] In step (708), the calculation unit (103) compares the determined pattern with the expected pattern. The calculation unit (103) may implement a selected ML model (e.g., 801a) to compare the determined pattern with the expected pattern. In one embodiment, the expected pattern may include a normal pattern. In one embodiment, the expected pattern may include both a normal pattern and an abnormal pattern. For example, when the received parameter is motor speed, the expected pattern may include abnormal changes in motor speed, and the selected ML model (e.g., 801a) determines that the motor parameter includes an abnormal pattern.

[0065] In step (709), the calculation unit (103) predicts breakage in the paper web based on comparison. When the determined pattern does not match the expected pattern, and when the expected pattern includes normal values, the selected ML model (e.g., 801a) predicts breakage in the paper web. When the determined pattern matches the expected pattern, and when the expected pattern includes outliers, the selected ML model (e.g., 801a) predicts breakage in the paper web. In one embodiment, the selected ML model (801a) provides the prediction result on the notification unit (104). Furthermore, the selected ML model (e.g., 801a) determines the root cause of breakage in the paper web based on the correlation between the outlier pattern in the parameters and the root cause of breakage in the paper web. Similarly, the selected ML model (e.g., 801a) estimates the breakage time in the paper web based on the correlation between the outlier pattern in the parameters and the estimated breakage time in the paper web.

[0066] Now for reference Figure 9 This shows the predicted breaks in the paper web for the marked parameters. For example... Figure 9 As shown, the label (601) is generated by an automatic labeling machine (304), and the label (901) is predicted using a selected ML model (e.g., 801a).

[0067] In one embodiment, the calculation unit (103) receives parameters after each break in the paper web. In rare cases, breakage in the paper web cannot be predicted, or corrective measures cannot be taken in time to prevent breakage. In such cases, a post-mortem analysis is performed. In the post-mortem analysis, the calculation unit (103) receives parameters after a break in the paper web occurs. Furthermore, the calculation unit provides the parameters to an automatic labeling machine (304) to generate labels including normal and abnormal patterns. In one embodiment, the automatic labeling machine can generate new labels including normal and abnormal patterns. In one embodiment, the parameters may belong to new behaviors, such as new grades or operating conditions, for which the selected ML model (801a) cannot predict breakage in the paper web. These new parameter changes or patterns are tagged and stored in a database (305).

[0068] In one embodiment, the computing unit (103) enables the selected ML model (e.g., 801a) to adaptively learn based on new or updated labels stored in the database (305). Now refer to Figure 8B The calculation unit (103) provides the marked data stored in the database (305) to the selected ML model (e.g., 801a). Furthermore, the calculation unit (103) identifies one or more new patterns from the marked data causing breakage in the paper web, updates the selected ML model (e.g., 801a) to accommodate the one or more new patterns, and configures the selected ML model (e.g., 801a) to predict breakage in the paper web based on the one or more new patterns. In one embodiment, the selected ML model (e.g., 801a) may be updated at regular time intervals (e.g., weekly, monthly, or semi-annually).

[0069] In one embodiment, the computing unit (103) performs a prescriptive analysis on multiple processing parameters and multiple operating parameters, as well as at least one parameter identified as the root cause of breakage in the paper web, to determine one or more actions to be performed to prevent breakage in the paper web. In one embodiment, a selected ML model (801a) or different ML models can be trained based on the root cause to provide one or more actions to be performed when breakage occurs in the paper web. For example, during the training of the ML model, when a motor speed parameter is identified as the root cause of breakage, a domain expert may have already performed actions such as shutting down the paper machine and scheduling maintenance activities for the motor. In real time, when breakage occurs due to an abnormality in motor speed, the ML model formulates a schedule for maintenance activities for the motor that caused the breakage.

[0070] Now for reference Figure 10 It demonstrates the use of a single model for paper web breakage prediction, a bidirectional long short-term memory (LSTM) encoder-decoder model based on multi-loss attention, and the root cause to predict breakage in paper web in online analysis. Figure 10 The graph shown is for lag data. In one embodiment, data processed using any feature engineering technique can be provided to the selected ML model (e.g., 801a). For example, a first difference, a second difference, various transformations, or any feature engineering method can be used with the ML model (e.g., 301a or 801a).

[0071] Figure 11 A graph illustrating the reconstruction error is shown. In one embodiment, the reconstruction error indicates the deviation between the determined pattern and the expected pattern. Figure 11 As shown, the star-shaped mark represents the marking abnormal area (reconstruction error) generated by the automatic labeling machine (304). The reconstruction error in the abnormal area is higher than that in the normal area.

[0072] Figure 12A and Figure 12B The learning curves of a multi-loss LSTM encoder-decoder model with a mean square error-based attention mechanism are shown. Figure 12A and Figure 12B As shown, the training and testing loss values ​​have decreased, and then overfitting begins. Figure 13 As shown Figure 12A and Figure 12B The diagram shown is a deep learning architecture diagram of the training graph.

[0073] In one embodiment, the computing unit (103) transfers knowledge from a selected ML model (801a) to a new ML model for predicting breakage in the paper web, determining the root cause of breakage in the paper web, and estimating the breakage time of the paper web in a new paper machine. For example, a paper mill may include two or more paper machines. The selected ML model (801a) is implemented for a first paper machine. Instead of implementing a new ML model (e.g., 801b) for a second machine, the ML model (e.g., 801a) can be transferred to the ML model (801b). Furthermore, the ML model (801b) can be trained with a limited dataset based on parameters provided to the second paper machine. In one embodiment, transfer learning can also be performed between paper mills. Figure 14 An exemplary block diagram is shown for transferring knowledge from a ML model (trained model) associated with a first paper machine to another ML model (untrained model) associated with a second paper machine.

[0074] In one embodiment, the present invention provides a framework for robustly labeling anomalous regions and efficiently predicting breaks in paper webs. When normal and anomalous regions are accurately labeled, the present invention reduces false positives while predicting breaks in the paper web and detecting only deviations that lead to breaks. The present invention reduces the manual requirements for labeling datasets while providing accuracy in generating labels. ML models are frequently retrained to incorporate new variations and parameters in the parameters. In one embodiment, the present invention automates the feasibility study of predicting breaks in paper webs and determining the root causes of breaks. In one embodiment, the present invention automates the process of labeling batch regions for predicting breaks in paper webs and labeling each parameter for predicting the root causes of breaks. In one embodiment, the present invention enables the integration of ML techniques to label regions, label parameters, predict breaks, and predict the root causes of breaks.

[0075] This written description uses examples to illustrate the subject matter, including best practices, and also enables any person skilled in the art to make and use the subject matter. The patentable scope of this subject matter is defined by the claims, and may include other examples that would occur to a person skilled in the art. These other embodiments are intended to be within the scope of the claims if they have structural elements that are not different from the literal language of the claims, or if they include equivalent structural elements that are not substantially different from the literal language of the claims.

[0076] Figure label:

[0077] 100 - Framework

[0078] 101 - Paper Machine

[0079] 102 - Data Processing Architecture

[0080] 103 - Computing Unit

[0081] 104 - Notification Unit

[0082] 301a-301n - ML Model (Offline Analysis)

[0083] 302 - Model Selector (Offline Analysis)

[0084] 303 - Memory

[0085] 304 - Automatic Labeling Machine

[0086] 305 - Database

[0087] 306a-306d - Abnormal

[0088] 307 - Fault Point

[0089] 401 - Exception Marker

[0090] 601 - Anomaly Labeling (Database Reduction)

[0091] 801a-801n - ML Model (Online Analysis)

[0092] 802 - Model Selector (Online Analysis)

[0093] 901 - Prediction Marker

Claims

1. A method for marking parameters associated with a paper machine (101) to predict breakage in the paper web within the paper machine (101), wherein the parameters include a plurality of process parameters and a plurality of operating parameters, wherein the method is performed by a computing unit (103), the method comprising: Simulation parameters related to the paper machine (101) are provided to multiple machine learning models, including normal mode and abnormal mode, the abnormal mode causing breakage in the paper web; Configure the plurality of machine learning models to label the simulation parameters as normal and abnormal modes, wherein the abnormal mode is close to the timestamp of the break in the paper web; Receive output from each of the plurality of machine learning models, wherein the output indicates a label including the normal mode and the abnormal mode; Based on one or more performance metrics and the outputs of the plurality of machine learning models, a machine learning model is selected from the plurality of machine learning models, and one or more model parameters of the selected model are stored in the memory (303) of the computing unit (103); Multiple details of the selected model are provided to an automated labeling machine independent of the machine learning model, and the automated labeling machine marks historical parameters as normal and abnormal modes, the historical parameters including at least one of normal and abnormal modes, wherein the labels generated by the automated labeling machine (304) are stored as label data in a database (305). The marked data is used to predict breaks in the paper web in real time, and Perform one or more actions on the paper machine to control it to avoid predicted breakage in the paper web.

2. The method of claim 1, wherein the simulation parameters and the historical parameters are received in multiple batches, wherein each batch includes the multiple process parameters and the multiple operating parameters, the multiple process parameters and the multiple operating parameters being simulated or measured between the time of restarting the paper machine (101) after a break in the paper web and the time to the subsequent break of the paper web in the paper machine (101).

3. The method of claim 1, further comprising labeling the simulation parameters and the historical parameters having the anomalous pattern with the root cause of the breakage in the paper web, wherein the root cause of the breakage in the paper web is included in the labeling data.

4. The method of claim 1, further comprising tagging the simulation parameters and the historical parameters having the anomalous pattern with the estimated time of paper web breakage, wherein the estimated time of paper web breakage is included in the tagging data.

5. The method of claim 1, wherein each batch is labeled as including the normal mode and the abnormal mode, based on independent analysis of each batch or by providing one of the historical batches to the automatic labeling machine.

6. A computing unit (103) comprising one or more processors and a memory (303), the computing unit being configured to perform the method according to any one of claims 1-5.

7. A method for predicting breakage in a paper web in a paper machine, wherein multiple sensors are used to monitor parameters associated with the paper machine (101), wherein the parameters include multiple process parameters and multiple operating parameters, wherein a paper web is formed in the paper machine (101) during papermaking, wherein the method is performed by a computing unit (103), the method comprising: The parameters are received from the plurality of sensors; Determine the pattern of how each parameter changes over time; Each of the determined patterns is compared with the corresponding expected pattern; and Based on the comparison, a tear in the paper web is predicted, wherein the corresponding expected pattern is generated using a machine learning model derived from multiple machine learning models, wherein training the multiple machine learning models includes: Labeled data including the parameters of normal and abnormal modes are provided to the plurality of machine learning models, wherein the timestamps associated with the outliers are close to the timestamps of the breaks in the paper web, wherein the labeled data is generated by an automatic labeling machine (304) independent of the plurality of machine learning models using simulation parameters and historical parameters associated with the paper machine (101) to label the historical parameters as normal and abnormal modes. Configure the plurality of machine learning models to detect patterns in the parameters and determine the detected patterns as at least one of normal and abnormal patterns; Configure the plurality of machine learning models to generate a desired pattern for each parameter based on the determined pattern, wherein each desired pattern includes at least one of a normal pattern and an abnormal pattern; Receive output from each of the plurality of machine learning models, wherein the output indicates a prediction of tearing in the paper web based on a comparison of the labeled data with a corresponding expected pattern; Based on the outputs of the plurality of machine learning models, a machine learning model is selected from the plurality of machine learning models; and Perform one or more actions on the paper machine to control it to avoid predicted breakage in the paper web.

8. The method of claim 7, wherein the plurality of machine learning models are further trained to: Based on the marked data and the abnormal patterns in the parameters, the time of paper web breakage is estimated, and the root cause of the breakage in the paper web is determined; and The correlation between generating the anomalous pattern and estimating the time of the paper web breakage, and the correlation between the anomalous pattern and the root cause of the breakage in the paper web.

9. The method of claim 7, wherein the selected model is further configured to: Abnormal patterns are identified from parameters received by one or more sensors by comparing each of the determined patterns with the corresponding expected pattern. The root cause of the breakage in the paper web and the estimated time of the breakage are determined based on the correlation between the abnormal pattern and the root cause of the breakage in the paper web, as well as the correlation between the abnormal pattern and the estimated time of the breakage in the paper web.

10. The method of claim 7, further comprising: Receive parameters after each break in the paper web; The parameters are provided to the automatic labeling machine (304) to generate labels including the normal mode and the abnormal mode; as well as The generated tags are stored as tag data in the database (305).

11. The method of claim 10, further comprising: The labeled data stored in the database (305) is provided to the selected machine learning model as feedback at defined time intervals; Identify one or more features from the marking data that caused the breakage in the paper web; and The selected machine learning model is updated to adapt to the one or more new patterns, and the selected machine learning model is configured to predict breaks in the paper web based on the one or more new patterns.

12. The method of claim 7, further comprising performing a prescribed analysis on the plurality of process parameters and the plurality of operating parameters, as well as the at least one parameter identified as the root cause of breakage in the paper web, to determine one or more actions to be performed to avoid the breakage in the paper web.

13. The method of claim 7, further comprising: One or more model parameters of the machine learning model are transferred to a new machine learning model to predict breakage in the paper web, determine the root cause of the breakage in the paper web, and estimate the time of breakage in the paper web in the new papermaking machine.

14. A computing unit (103) comprising one or more processors and a memory, the computing unit being configured to perform the method according to any one of claims 7-13.

Images