Adaptive-learning, auto-labeling method and system for predicting and diagnosing web breaks in paper machine

Patent metadata

Field	Value
Publication number	US-12460349-B2
Application number	US-202217993924-A
Country	US
Kind code	B2
Filing date	Nov 24, 2022
Priority date	May 26, 2020
Publication date	Nov 4, 2025
Grant date	Nov 4, 2025

Abstract

A system and method for labelling normal and abnormal regions in data related to a paper machine for web break prediction and labelling individual parameters for root cause analysis, using machine learning models, includes using the machine learning models in real-time to predict breaks in the paper web, analyzing root cause for the breaks in the paper web, and estimating a time to break. An auto-data-labeling framework helps in adaptive learning for autonomous model improvement of the deployed model, transfer learning, shortlisting parameters and automating feasibility study.

First claim

What is claimed is:

1 . A method of labelling parameters related to a paper machine to predict a break in paper web in the paper machine, wherein the parameters comprise a plurality of process parameters and a plurality of operational parameters, wherein the method is performed by a computing unit, the method comprising:
providing simulated parameters related to the paper to a plurality of machine learning models, the simulated parameters comprising normal patterns and abnormal patterns, the simulated parameters being known to have caused a break in the paper web; configuring the plurality of machine learning models to label the simulated parameters into normal patterns and abnormal patterns, wherein the abnormal patterns are proximate to a timestamp of the break in the paper web; receiving an output from each of the plurality of machine learning models, wherein the output is indicative of labels comprising the normal patterns and the abnormal patterns; selecting a machine learning model from the plurality of machine learning models based on one or more performance metrics and the output of the plurality of machine learning models, and storing one or more model parameters of the selected model in a memory of the computing unit; providing a plurality of details of the selected model to an auto-labeller independent of the machine learning model and labelling, by the auto-labelller, the historical parameters into the same normal patterns and the same abnormal patterns, the historical parameters comprising at least one of the normal patterns and the abnormal patterns, wherein the labels generated by the auto-labeller are stored as labelled data in a database; using the labelled data for predicting a break in the paper web in real-time; and performing one or more actions to the paper machine to control the paper machine to avoid the predicted break in the paper web.

2 . The method as claimed in claim 1 , wherein the simulated parameters and the historical parameters are received in a plurality of batches, wherein each batch comprises the plurality of process parameters and the plurality of operating parameters simulated or measured between a time of restarting the paper machine after a break in the paper web, to a time of a subsequent break of the paper web in the paper machine.

3 . The method as claimed in claim 2 , wherein each batch of the plurality of batches is labelled to comprise the normal patterns and the abnormal patterns based on providing similar batches to the auto-labeller, wherein the similar batches share at least one aspect with a provided batch of the plurality of batches.

4 . The method as claimed in claim 1 , further comprising labelling the simulated parameters and the historical parameters having the abnormal patterns with a root cause for the break in the paper web, wherein the root cause for the break in the paper web is included in the labelled data.

5 . The method as claimed in claim 1 , further comprising labelling the stimulated parameters and the historical parameters having the abnormal patterns with an estimated time to break the paper web, wherein the estimated time to break the paper web is included in the labelled data.

6 . The method as claimed in claim 1 , wherein each batch is labelled to comprise the normal patterns and the abnormal patterns based on independently analyzing each batch.

7 . The method as claimed in claim 1 , wherein the auto-labeller independent of the machine learning model is a separate, second machine learning model from the plurality of machine learning models.

8 . The method as claimed in claim 1 , wherein the simulated parameters are associated with a defined time interval between a first break in the paper web and a second break in the paper web.

9 . The method as claimed in claim 1 , further comprising:
receiving the parameters after each break in the paper web; providing the parameters to an auto-labeller for generating labels comprising the same normal patterns and the same abnormal patterns; storing the generated labels as labelled data in a database; providing the labelled data stored in the database to the selected machine learning model as a feedback at defined intervals of time; identifying one or more new patterns from the labelled data that caused the break in the paper web; and updating the selected machine learning model to adapt to the one or more new patterns and configure the selected machine learning model to predict the break in the paper web based on the one or more new patterns.

10 . The method as claimed in claim 1 , further comprising:
transferring knowledge of the selected machine learning model to a new machine learning model for predicting a break in a paper web of a second paper machine; and using the new machine learning model to predict the break in the paper web of the second paper machine in real-time.

11 . A method of predicting a break in a paper web in a paper machine, wherein a plurality of sensors are used to monitor parameters related to the paper machine, wherein the parameters comprise a plurality of process parameters and a plurality of operating parameters, wherein a paper web is formed in the paper machine during manufacturing of paper, wherein the method is performed by a computing unit, the method comprising:
receiving the parameters from the plurality of sensors; determining a pattern in a variation of each parameter, with respect to time; comparing the determined pattern with a corresponding expected pattern; and predicting a break in the paper web based on the comparison; wherein the corresponding expected pattern is generated using a selected trained machine learning model from a plurality of machine learning models, wherein training of the plurality of machine learning models comprises: providing labelled data comprising normal patterns and abnormal patterns of the parameters, to the plurality of machine learning models, wherein a timestamp associated with the abnormal values is proximate to a timestamp of the break in the paper web, wherein the labelled data is generated using an auto-labeller independent of the plurality of machine learning models using simulated parameters and historical parameters related to the paper machine to label the historical parameters into the same normal patterns and the same abnormal patterns; configuring the plurality of machine learning models to detect patterns in the parameters and determine the detected patterns into at least one of normal patterns and abnormal patterns; configuring the plurality of machine learning models to generate the expected pattern for each parameter based on the determined patterns, wherein each expected pattern comprises at least one of, normal pattern and abnormal pattern; receiving an output from each of the plurality of machine learning models, wherein the output is indicative of a prediction of a break in the paper web based on a comparison of the labelled data with corresponding expected patterns; selecting a machine learning model from the plurality of machine learning models, based on the output of the plurality machine learning models, as the selected trained machine learning model; and performing one or more actions to the paper machine to control the paper machine to avoid the predicted break in the paper web.

12 . The method as claimed in claim 11 , wherein each of the plurality of machine learning models is further trained to:
estimate a time to break the paper web and determine a root cause that causes a break in the paper web based on abnormal patterns in the parameters and the labelled data; and generate an association between the abnormal patt

Assignees

• Abb Schweiz Ag

Inventors

• Ramu Vadthyavath
• Patil Dinesh
• Kubal Nandkishor

Classifications

• D21G9/0054

  details of algorithms or programs · CPC title

• G06F30/27

  using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model · CPC title

• G06V30/194

  References adjustable by an adaptive method, e.g. learning · CPC title

• G06F18/2185

  the supervisor being an automated module, e.g. intelligent oracle · CPC title

• G06F18/2148

  characterised by the process organisation or structure, e.g. boosting cascade · CPC title